Thesis Giacomo Accorto Master

Università degli Studi di Milano
FACOLTÀ DI SCIENZE E TECNOLOGIE

Corso di Laurea Magistrale in Fisica
Tesi di Laurea Magistrale
A Constrained-Minimization Method
for the Solution of the
Inverse Kohn-Sham Problem in Nuclei
Relatore: Candidato:
Prof. Gianluca Colò Giacomo Accorto
Matricola 884296
Correlatore:
Dott. Xavier Roca-Maza
Anno Accademico 2017–2018

Contents
1 Nuclear DFT 3
1.1 Nuclear Phenomenology and QM . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Nuclear forces and Bulk Properties of Nuclei . . . . . . . . . . . . 3
1.1.2 Basic Concepts of Quantum Mechanics . . . . . . . . . . . . . . . 6
1.1.3 Nuclear Shell Model . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Density Functional Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.1 Hohenberg-Kohn Theorems . . . . . . . . . . . . . . . . . . . . . 12
1.3 Levy-Lieb Constrained-Search Formulation of DFT . . . . . . . . . . . . 14
1.3.1 N-Representable and v-Representable Densities . . . . . . . . . 14
1.3.2 Levy-Lieb Constrained-Search Formulation . . . . . . . . . . . . 15
1.4 KSDFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 IKS Problem in Nuclear DFT 19

2.1 Forward and Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Regularization techniques for Ill-posed Problems . . . . . . . . . . . . . 20
2.2.1 Tikhonov Regularization . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 IKS Problem in Nuclear DFT . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1 Comparing the Inverse and the Forward Problem . . . . . . . . . 22
2.3.2 Inversion methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 CV Inversion Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.1 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 CV Method in Spherical Symmetry . . . . . . . . . . . . . . . . . . . . . . 32
2.5.1 The Spherical Hamiltonian . . . . . . . . . . . . . . . . . . . . . . 32
2.5.2 The Spherical Form of the Cost Functional . . . . . . . . . . . . . 34
2.5.3 Spherical Bondary Conditions . . . . . . . . . . . . . . . . . . . . 38
2.6 Parametrizations of Experimental Nuclear Densities . . . . . . . . . . . 39
3 Implementation and Inversion Results 43

3.1 Implementation of the Constrained Variational Method . . . . . . . . . 43
3.2 Unidimensional Numerical Tests . . . . . . . . . . . . . . . . . . . . . . . 45
3.2.1 Harmonic Oscillator Density . . . . . . . . . . . . . . . . . . . . . 45
3.2.2 Morse Oscillator Density . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Three-dimensional Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4 KS Potential from Nuclear Densities . . . . . . . . . . . . . . . . . . . . . 59
3.4.1 The Helium Nucleus . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.4.2 The Oxygen Nucleus . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4.3 The Calcium Nucleus . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.4.4 Lead Nucleus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
i
ii CONTENTS
A The Method of Lagrange Multipliers 81
B Density Operators and Scalar/Vector Densities 83

B.0.1 Density Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
B.0.2 Reduced Density Matrices . . . . . . . . . . . . . . . . . . . . . . . 84
B.0.3 Spinless density matrices . . . . . . . . . . . . . . . . . . . . . . . 84
B.0.4 Scalar and Vector Densities . . . . . . . . . . . . . . . . . . . . . . 85
C Time-Independent Systems 87
D Deconvolution from Proton Charge Density to Proton Density 89
E Discretization Methods 93
E.0.1 Discretized Derivative . . . . . . . . . . . . . . . . . . . . . . . . . 93
E.0.2 Discretized Integration . . . . . . . . . . . . . . . . . . . . . . . . . 93
E.0.3 Linear Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
E.0.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Bibliography 97
List of Figures
1.1 The nuclear density, in the inner part of nuclei, is in good approximation
constant and independent of the mass number. . . . . . . . . . . . . . . . . . 4
1.2 The binding energy per nucleon is, in first approximation, a constant function. 5
1.3 The nuclear shell structure; on the left the oscillator shells, on the right the
spin-orbit interaction produces the j -splitting and predict the correct magic
numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1 The Kohn-Sham orbital for a single particle, subject to a harmonic oscillator
trap, is Gaussian and concide with the first eigenfunction of the harmonic
oscillator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 The Kohn-Sham potential as it results from a non-scaled one-orbital harmonic
oscillator inversion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3 The Kohn-Sham potential, computed through the one-orbital anlaytic for-
mula (3.13), compared with the expected potential. . . . . . . . . . . . . . . . 48
3.4 The Kohn-Sham potential as obtained from a scaled one-orbital harmonic
oscillator inversion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5 The first Kohn-Sham orbitals resulting from a two-orbital inversion in absence
of a regularization scheme, compared with the same orbital calculated using
the Tikhonov regularization. The number of inflexion points diminishes. . . . 50
3.6 The second Kohn-Sham orbitals resulting from a two-orbital inversion in ab-
sence of a regularization scheme, compared with the same orbital calculated
using the Tikhonov regularization. The asymmetries disappear. . . . . . . . . 50
3.7 The Kohn-Sham orbitals resulting from a three-orbital inversion (solid lines)
compared with the first three eigenfunction of the harmonic oscillator (dashed
lines). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.8 The Harmonic Kohn-Sham potential in a three-orbital system, as recovered
from the Kohn-Sham orbitals shown in figure 3.7. . . . . . . . . . . . . . . . . 51
3.9 The computational time is a steep function of the number of filled orbital
(degrees of freedom of the problem) in the harmonic oscillator inversion. A
2
qualitative comparison of such function with e n /60 is also illustrated in the
figure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.10 The Morse Oscillator for D = 10 and α = 1/2. . . . . . . . . . . . . . . . . . . . 53
3.11 The Kohn-Sham potential from the density of seven-orbital in a Morse oscillator
trap. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.12 The density of a system composed by twenty particles, distributed in the four
lowest-energy orbitals, in an isotropic harmonic oscillator trap. . . . . . . . . 57
iii
iv LIST OF FIGURES
3.13 The Kohn-Sham orbitals that are obtained by minimizing the objective function
highly resemble the 1s, 1p, 1d , and 2s eigenfunctions (respectively from left to
right, top to bottom) of the harmonic oscillator. . . . . . . . . . . . . . . . . . 57
3.14 The Kohn-Sham potential of a system composed by twenty particles, distributed
in the four lowest-energy orbitals, versus the expected result. The potentials
are shifted to provide a better qualitative comparison. . . . . . . . . . . . . . 58
3.15 The 4 He proton density as obtained via a deconvolution from the proton charge
densy provided by De Vries [1]. . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.16 The Kohn-Sham potential deduced from the 4 He proton density. . . . . . . . 62
3.17 The Kohn-Sham potential, deduced from the 4 He proton density, presents a
non-physical parabolic growth at large r . . . . . . . . . . . . . . . . . . . . . 62
3.18 The 16 O proton density as obtained via a deconvolution from the proton charge
density provided by De Vries [1] with a 3pF and a 12SoG parametrizations. . . 64
3.19 The 16 O proton density as obtained via a deconvolution from the proton charge
density provided by De Vries [1] with a 3pF and a 12SoG parametrizations.
Logarithmic scale is used to highlight the functional behaviour of the tails. . . 65
3.20 A comparison between the Kohn-Sham potential obtained by 16 O proton den-
sity (12SoG) through the constrained-variational method and a Woods-Saxon
potential. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.21 Comparison between the Kohn-Sham potential obtained by 16 O proton density
(3pF) through the constrained-variational method and a Woods-Saxon potential. 66
3.22 The 40 Ca proton density as obtained via a deconvolution from the proton
charge density provided by De Vries [1] with a 3pF and a 12SoG parametriza-
tions. The density of twenty particles in an isotropic harmonic oscillator trap is
also depicted. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.23 The 40 Ca proton density as obtained via a deconvolution from the proton
charge density provided by De Vries [1] with a 3pF and a 12SoG parametriza-
tions. The density of twenty particles in an isotropic harmonic oscillator trap is
also depicted. Logarithmic scale is used to highlight the functional behaviour
of the tails. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.24 Comparison between Kohn-Sham potential deduced by the experimental 12SoG
proton density of 40 Ca and that obtained from the density of twenty particles
in an isotropic harmonic oscillator trap. . . . . . . . . . . . . . . . . . . . . . 69
3.25 Comparison between the Kohn-Sham potential obtained by 40 Ca proton den-
sity (3pF) through the constrained-variational method and a Woods-Saxon
potential. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.26 The neutron Lead density provided by Zenihiro [2] compared with the density
obtained by solving a direct problem within Hartree-Fock approximation, in
which a Skyrme force, SkP [3], is assumed. . . . . . . . . . . . . . . . . . . . . 71
3.27 The neutron Lead density provided by Zenihiro [2] compared with the density
obtained by solving a direct problem within Hartree-Fock approximation, in
which a Skyrme force, SkP [3], is assumed. Logarithmic scale is used to better
analyse the tails of the densities. . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.28 The Kohn-Sham potential as deduced from the experimental neutron density,
compared with a Woods-Saxon potential. . . . . . . . . . . . . . . . . . . . . 72
3.29 The Kohn-Sham potential as deduced from the experimental neutron density,
with a corrected Woods-Saxon tail. . . . . . . . . . . . . . . . . . . . . . . . . 72
3.30 The Kohn-Sham potentials deduced from the Hartree-Fock theoretical and from
the experimental neutron densities, compared with the potential assumed in
the Hartree-Fock calculations of the density. . . . . . . . . . . . . . . . . . . . 75
LIST OF FIGURES v
3.31 The Kohn-Sham potential deduced from experimental neutron density through
two different inversion schemes, the CV method and the vLB method, compared
with the potential assumed in the Hartree-Fock calculations of the density. . . 76
3.32 The density that is obtained by solving the Kohn-Sham direct problem using
the Kohn-Sham potential deduced by experimental neutron density, compared
with the experimental neutron density. . . . . . . . . . . . . . . . . . . . . . 77
3.33 The density obtained by solving the Kohn-Sham direct problem using the Kohn-
Sham potential deduced by experimental neutron density compare with the
experimental neutron density. Logarithmic scale is used to analyse the tails of
the densities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.34 The relative spread between the density obtained by solving the Kohn-Sham di-
rect problem using the Kohn-Sham potential deduced by experimental neutron
density and the experimental neutron density. . . . . . . . . . . . . . . . . . 78
Introduction
Density functional theory (DFT) represents one of the most successful theoretical
approaches to the study of both electronic [4] and nuclear structure [5,6]. In particular,
DFT is the only theory that allows exploring systematically both medium-mass and
heavy nuclei, and it is by far the model that permits to study the highest number of
elements of the nuclear chart (all of them, at least in principle). The theory states
a one-to-one correspondence between the correct ground state density n(~ r ) of a
fermion system and the minimum value of an energy density functional (EDF) E [n].
Also, given the ground state density of a specific many-body system, one has access
to its one-body ground state properties, such as charge and neutron radii, electric
and magnetic moments, and so on [7].
The original formulation of density functional theory, provided by P. Hohenberg
and W. Kohn in 1964 [8], is based on proof-by-contradiction theorems about the
existence of an exact energy density functional, but do not point to any constructive
method to get its form. Instead, the theory can be exploited as a rule, for precise
calculations, in the Kohn-Sham scheme [9], that consists in a clever mathematical
reorganization of the energy density functional. An auxiliary, non-interacting, Kohn-
Sham system is uniquely defined once its density, that can be written in terms of
single-particle orbitals, is set equal to that of the original system. When the equality
is established, the energy density functional of the fictitious system coincides with
that of the original one, as well as the ground state properties related to one-body
operators. In such framework, one is able to write down the terms composing the
EDF, namely a non-interacting kinetic energy term and a contribution given by an
effective Kohn-Sham potential, coupled to the density. Although the formalism highly
recalls that typical of mean-field methods, such as Hartree-Fock, it is remarkable that
Kohn-Sham DFT provides, at least in principle, a fully exact formalism. A given level
of approximation, such as local density approximation (LDA), must be applied only
for practical calculations, in order to establish the form of the Kohn-Sham potential.
The present thesis addresses the solution of the inverse problem [10, 11] in Kohn-
Sham density functional theory (KSDFT), in relation to nuclear systems. The problem
consists in deducing the form of the Kohn-Sham potential once the density function
of a nucleus is given. The aim is indeed pioneering in the field of Nuclear Physics,
where the knowledge of the Kohn-Sham potential could provide an useful benchmark
to the state-of-art approximate energy density functionals. In fact, the literature
on the argument is almost entirely due to works that have been done in the field
of Atomic and Condensed Matter Physics, or in Quantum Chemistry. The Milan
Nuclear Physics research group has begun being interested in the inverse Kohn-Sham
problem in 2015. Since then, three bachelor’s theses, [12–14], have been devoted to
the argument.
Many methods, seemingly independent of one another, have been developed
over the last thirty years to solve the inverse Kohn-Sham problem in DFT. These
methods emerge from different ways of formulating the inverse problem. Some
of them are based on the optimization of a functional [15, 16], while others are
iterative [17–21]. The purpose of the present master’s thesis is to try developing a
constrained-minimization inversion algorithm, independent of, and more formal
than, the one that has been used in the previous works. Our algorithm is inspired by
the so-called Constrained-Variational method, that has been proposed in a recent
paper by S. Jensen and A. Wasserman [22]; however, the article provides tests with
analytic unidimensional densities, only. It is our scope to generalize such method
to render it applicable to nuclear densities. In particular, we are interested to study
of spherical symmetric magic nuclei, in which specific nuclear effects, such as the
pairing, are negligible.
Chapter one of the thesis is devoted to a general review of nuclear phenomenology
and to the exposition of the theoretical details concerning density functional theory.
Chapter two begins discussing the argument of inverse problems and how they
require a more careful treatment than direct ones. In particular, we deepen the
differences in the implementation of the the direct and inverse Kohn-Sham prob-
lem. Afterwards, we present the details of our inversion method, its theoretical
assumptions, and its application to spherical symmetric systems. Finally, we discuss
the methods through which the input data of the inverse problem, that is nuclear
densities, are provided by the experimental literature.
The third chapter consists of a discussion of the numerical methods and libraries
we used to implement the inverse problem solution and of the analysis of the results
of the numerical tests we have benchmarked our software against. First, a set of uni-
dimensional analytic tests, then a three-dimensional isotropic analytic test. The last
section discusses the results we have obtained in applying the density-to-potential in-
version method to the density of four magic nuclei: 4 He, 16 O, 40 Ca and 208 Pb. Finally,
our conclusions will be drawn.
Chapter 1
Nuclear Density Functional

Theory
1.1 Introduction to Nuclear Phenomenology and to Ba-

sic Concepts of Quantum Mechanics
1.1.1 Nuclear forces and Bulk Properties of Nuclei
Nuclei are self-bound systems, made by two types of fermions: protons and
neutrons. Together, they are called nucleons. The forces that keep them bound are the
nuclear forces, a general set of interactions comprising nucleon-nucleon, pionic and
heavy mesons interactions. In general, a standard feature of the interactions of the
many-body nuclear problem is that the forces acting between each pair of fermions
involve the interplay of all hadrons, at least to some extent. In fact, depending on the
scale of energy at which phenomena of interest happen, there is a different probability
of producing a variety of strongly interacting particles. As we will remark below, the
distance between nucleons in nuclei is large enough for the strong repulsive core of
the interaction not to be felt at full by nucleons. Instead, nucleons mainly experience
the softer tail of the interaction. Consider that the energy scale for removing a nucleon
from a nucleus is S (n,p) ≈ 10 MeV, while the kinetic energy of nucleons in nuclei is
T ≈ 40 MeV; in contrast, the rest mass of the nucleons is m (n,p) c 2 ≈ 1 GeV, and
that of the lightest existing hadron, the pion, is m π c 2 ≈ 137 MeV. A reasonable first
approximation is to consider nuclei as being composed by non-relativistic nucleons.
The exchange of mesons between neutrons or protons can be taken into account
through the action of effective forces. Thereby, the full complexity of the nuclear
forces does not come into play.
In addition to the nuclear interactions, the influence of other weaker forces
cannot be neglected. For instance, protons indeed feel, and generate, electromagnetic
fields. Also, the presence of weak interactions manifests itself in processes such as
β-decays. Albeit such minor forces are not fundamental for establishing the main
bulk properties of the nuclear structure, they play an important role in determining
the stability of bound states.
Because of the short range of the nuclear forces, all nuclei are characterized, in
3
4 CHAPTER 1. NUCLEAR DFT
Figure 1.1: The nuclear density, in the inner part of nuclei, is in good approximation constant
and independent of the mass number.
their interior part, by an approximately constant density (see figure 1.1) with value
n(0) = 0.16 fm−3 (1.1)
and by a volume that grows qualitatively in proportion with the mass number A. The
density typically decay to zero over a distance of 2 fm; therefore the surface thickness
of nuclei is rather small if compared to their radial extent R, defined as the distance
at which the density has become half of its interior value n(0). An empirical value for
this quantity is given by
1
R = r0 A 3 ; (1.2)
³ ´1
3 3
electron elastic scattering data determine r 0 = 4πn(0) ≈ 1.15 fm.
The available data about nuclear matter distributions almost entirely consists
of proton densities [1]. The neutron density of few isotopes of lead and tin have
been obtained via proton scattering [2]. In heavy nuclei, the excess of neutrons with
respect to protons gives rise to differences between the densities of the two types of
nucleons.
Another important feature of nuclei is the value of the mean free path for the
collision between the constituent nucleons. Many are the evidences that the mean
free path in a nucleus is large in comparison with the nuclear size. Such long mean
free path of the nucleons entails that the interactions, in first approximation, result in
a smooth average potential, in which the particles move almost independently one
of each other.
A simple way to understand the main trends characterizing the nuclear structure
is to analyse the well-known semi-empirical expression provided by Weizsäcker in
1935 for the total nuclear binding energy. Such quantity is defined as the difference
between the observed total nuclear energy in the ground state and the rest masses of
the separeted nucleons. Based on the rough features we have discussed above, it is
1.1. NUCLEAR PHENOMENOLOGY AND QM 5
Figure 1.2: The binding energy per nucleon is, in first approximation, a constant function.
possible to write down a simple mass formula for the binding energy:
2 1 (N − Z )2 3 Z 2 e 2
B = b vol A − b surf A 3 − b sym − (1.3)
2 A 5 Rc
• b vol A is the main, always present, volume energy term. It represents the limit
of the binding energy for large A, N = Z nuclei, in absence of Coulomb forces.
The bulk binding energy of a nucleus is proportional to A and not to A(A − 1)/2
because of the short range of the interaction: each nucleon interacts with its
neighbours, only.
2
• b surf A 3 is the surface energy, a contribution typical of finite systems, that
reflecs the fact that particles are less bound at the surface, for they have less
neighbours.
• The third term takes into account the fact that symmetric configurations of
nuclei are energetically preferred. The Pauli principle enforces protons and
neutrons to occupy energy levels with increasing energy. An inbalance of one
type of fermion with respect to the other implies higher energy levels to be
occupied.
• The last term represents the Coulomb energy of a uniform sphere with radius
R c and it is responsible for the slight, gradual, decrease of the binding energy
per particle that can be observed in heavy nuclei. Also, it causes the excess of
neutrons in heavy stable nuclear systems.
Figure 1.2 clearly shows the above trends. Although the Weizsäcker formula is
subject to improvements by the addition of extra terms, such as a pairing, it does
remain an empirical, qualitative tool, that completely neglects quantum effects.
However, it is a noteworthy example of how different effects interplay in nuclear

systems and it accounts for their relative relevance.
From the observed binding energies, one can deduce the order of magnitude
of the average potential energy that a single nucleon feels in a nucleus. The energy
that is required to remove a neutron from a nucleus, namely the neutron separation
energy, is
B(N , Z ) − B(N − 1, Z ) ≈ 10 MeV ≈ −(Vn + ²F ). (1.4)
Since ²F ≈ 40 MeV, we expect such average potential, felt by a single nucleon, to be
characterized by a depth of Vn ≈ −50 MeV.
Just as in the electronic structure of atoms, the degeneracy of single-particle or-
bitals entails marked discontinuities, namely shell effects, in many nuclear properties.
Therefore, nuclear configurations based on energy levels of nucleons in nuclei can be
developed. Those constitute the nuclear shell structure.
1.1.2 Basic Concepts of Quantum Mechanics

Before explaining in detail how to develop the Nuclear Shell Model, let us recall
some basic concepts of quantum mechanics. This will be useful in order to set the
notation, a proper enviroment for the theoretical models exposed in the rest of the
chapter and to define some relevant quantities.
The investigation of nuclei in their ground state concerns the solution of the time-
independent Schrödinger equation. In fact, time-reversal symmetry holds in many
cases of interests, namely in the study of even-even nuclei. For an A-particle system
in the non-relativistic approximation, the Schrödinger equation has the structure of
an eigenvalue problem,
Ĥ Ψ = E Ψ, (1.5)
where E is the energy of the nucleus, Ψ = Ψ(~ x N ) is the total wave function of
x 1 , . . . ,~
the system, Ĥ is the Hamiltonian operator,
A ~2 A
∇2i +
X X X
Ĥ = T̂ + V̂ + Ŵ = − v(~
xi ) + w(~
x i ,~
x j ), (1.6)
i =1 2m i i =1 i 6= j
and the coordinates ~ x i comprise both the space coordinates ~ r i and the spin coor-
dinates σi . The form of such Hamiltonian is pretty general. The idea is to provide
an interdisciplinary notation, valid for the study of different fermionic system. It is
clear that, for instance, in the nuclear case no external potential is present. On the
other hand, as we have discussed above, the relevance of three-body interactions is
quantitative. However, equation (1.6) is adequate to the purpose of presenting, in the
next sections, the background theory that is fundamental to understand the scopes
of the work of thesis.
Boundary conditions are required to make the problem solvable. In particular,
the wave function Ψ must be well-behaved, that is smooth, everywhere. For the case
of a finite nucleus, the wave function must decay to zero at infinity; if one investigates
infinite nuclear matter, it must respect periodic boundary conditions. Because of the
fermionic nature of the nucleons, the wave function Ψ must be antisymmetric with
respect to the interchange of two particles.
Consider the set of the eigenvectors Ψk of Ĥ , associated to the energy eigenvalues
E k . It follows from basic algebraic theorems that the set {Ψk }k is complete, and its
elements can be chosen as orthogonal and normalized. Among all of the eigenvectors,
there is one that is related to the lowest energy eigenvalue E 0 : the ground state wave
function Ψ0 .
Each property of the system in some state can be calculated via an integration
over 3N spatial coordinates and a summation over N spin coordinates. For instance,
the expectation value of an observable, linked to a Hermitian linear operator Â, is
given by
x Ψ∗ ÂΨ 〈Ψ| Â|Ψ〉
R
d~
〈 Â〉 = RΩ = . (1.7)
Ω d~x Ψ∗ Ψ 〈Ψ|Ψ〉
One can read the equation above in the sense that 〈 Â〉 is a functional of the state Ψ,
〈 Â〉 = A[Ψ]. (1.8)
Consider the measurement of the energy expectation value
〈Ψ| Ĥ |Ψ〉
E [Ψ] = ; (1.9)
〈Ψ|Ψ〉
each measurement of the energy gives as a result a linear combination of the eigen-
values E k of Ĥ . The ground state energy is the global lower bound of the energy, as
computed from any Ψ. A full minimization of the functional E [Ψ] with respect to all
the allowed A-nucleons wave functions is therefore the ground state energy
E [Ψ0 ] = E 0 = min E [Ψ]. (1.10)
Ψ
In fact, it is indeed always possible to decompose any state Ψ on the complete set of
the eigenstates of the Hamiltonian
N
Ψ= C k Ψk .
X
(1.11)
k=1
This decomposition can be explicitely inserted into the energy functional
PN
|C |2 E k
k=1 k
E [Ψ] = P N
. (1.12)
k=1 k
|C |2
Since it is always possible to establish an ordering of the energy spectrum as E 0 ≤
E 1 ≤ E 2 ≤ · · · ≤ E [Ψ], E [Ψ] will be greater than the eigenvalue E 0 and the minimum
will be reached if and only if Ψ = C 0 Ψ0 . Every eigenstate of the Schrödinger equation
is a local extremum of E [Ψ], and it is therefore possible to replace the Schrödinger
equation itself with the variational principle
δE [Ψ] = 0. (1.13)
It is possible to reformulate the variational principle in such a way that the final state
Ψ will be automatically normalized; one extremizes the quantity 〈Ψ| Ĥ |Ψ〉 subject
to the constraint 〈Ψ|Ψ〉 = 1. According to the Lagrange multipliers method (see
appendix A) this can be done by imposing
δ[〈Ψ| Ĥ |Ψ〉 − E 〈Ψ|Ψ〉] = 0. (1.14)
The above procedure prescibes a method for moving from the piece of infor-
mation about N and v(~ r ), to the ground state wave function Ψ0 , and hence to the
ground-state properties of interest. It is very remarkable that the kinetic term and
the inter-particle potential, depending only on the numeber N of nucleons, are not
mentioned at all by the procedure: they are universal, in a sense that will be deepened
below, by means of the Hohenberg and Kohn theorems. We highlight again that the
energy E [Ψ] is a functional of N and v(~ r ), alone.
1.1.3 Nuclear Shell Model

It is easy to understand that finding an exact solution of the Schrödinger equation
for finite systems, such as nuclei, often reveals to be highly difficult. Luckily, most
of the properties of finite nuclei can be well predicted within models that make use
of approximate single-particle potentials. We remark again that the usage of single-
particle potentials is thoretically justified by the long mean free path of nucleons
in nuclei. The simplest forms of the single-particle potential one can consider are
the infinite square-well potential and the harmonic potential, in spherical symmetry.
Those provide the starting point for introducing the single-particle Nuclear Shell
Model at its roughest approximation.
In the former case, the eigenfunctions of the Schrödinger equation, respecting
proper boundary conditions, read
Ψnl ml (~
r ) = Nnl j l (kr )Yl ml (θ, φ), (1.15)
where Nnl is a normalization constant, j l are Bessel functions and Yl ml are the well-
known spherical harmonics. Such eigenfunctions must vanish at the boundary of
the well, which means that k nl R = X nl must be the n th zero of the l th Bessel function.
Their corresponding energy eigenvalues are
~2 knl
2
²nl = − v0. (1.16)
2m
The ordering of those eigenvalues produces the energy levels.
A second solvable single-particle potential is the three-dimentional isotropic
harmonic oscillator
1
v(r ) = −v 0 + mω2 r 2 . (1.17)
2
Consider an Ansatz of the eigenfunctions in the form
u nl
Ψnl ml (~
r)= Yl ml (θ, φ). (1.18)
r
When those are inserted into the Schrödinger equation, they give rise to a radial
differential equation
h ~2 d 2 ~2 l (l + 1) i
− + v(r ) + − (²nl + v 0 ) , (1.19)
2m d r 2 2m r2
that is solved by
q2 l+ 1
u nl (r ) = Nnl q l +1 e − 2 L n−12 (q 2 ) (1.20)
Γ(a + p + 1) e z d p a+p −z
L ap (z) = (z e ) (1.21)
Γ(p + 1) z a d z p
2(n − 1)!
Nnl = ; (1.22)
b[Γ(n + l + 21 )]3
where q = br , b 2 = mω
~
and L ap (z) are the generalized Laguerre polynomials. The
spectrum of energies is obviously that of a harmonic oscillator
3
²nl = ~ω(N + ) − v 0 , (1.23)
2
N = 2(n − 1) + l = 0, 1, 2, . . . , (1.24)
and produce the oscillator shells, each spaced from the other by ~ω. A proper semi-
1
empirical value for the level spacing is ~ω ≈ 41/A 3 MeV. Such value is a rough
estimate thought in order to fit to mean-square radii of nuclei.
Albeit the true nuclear single-particle potential should definitely have a finite
depth, the two previous approximations yield to a rather acceptable prediction of
low-lying levels, those with higher binding energy. The higher is the quantum number
l , the closer the levels are to the edge of the true well. The square-well potential in
general predicts stronger bindings than the harmonic oscillator potential.
Having defined such shell configuration of nucleons, one finds that a partic-
ular stability is reached for closed-shell configurations. In particular, the highest
stability is found in correspondence of numbers of protons or neutrons equal to
2, 8, 20, 28, 50, 82, 126, . . . (see figure 1.3). Nuclei with such number of protons or neu-
trons are called magic or doubly magic nuclei and they are characterized by spherical
symmetry.
Both the harmonic oscillator and the square well shell model correctly predict
only the first three shell clusures. In 1955 M. G. Meyer and J. H. D. Jensen [23]
proposed to consider a single-particle spin-orbit interaction term
H 0 (r ) = −α(r )L̂ · Ŝ, (1.25)
where α(r ) is a coupling costant, while L̂ and Ŝ are the angular momentum and the
spin operators, respectively. Such term is much more relevant that in the atomic
case, and its presence leads to the appearance of a splitting of levels characterized
by different total angular momentum. In the case of nuclear bindings the spin-orbit
levels with higher total angular momentum j at given l are found to lie lower in
energy; To account for this evidence, the spin-orbit interaction must be attractive,
that is the coupling constant must be positive. The best choice of quantum numbers
to treat the spin-orbit term is indeed the coupled basis |nl 12 j m j 〉. The first order
splitting is equal to
2l + 1
²l − 1 − ²l + 1 = αnl >0 (1.26)
2
Z
2 2
2
αnl = d r u nl (r )α(r ). (1.27)
The addition of this term allows to predict the correct magic numbers, since it can
happen that high j and l levels may be pushed from one shell to the lower one.
Many developments of the shell model have been performed, such as the addition
of new terms to the potential, or the choice of a different single-particle potential. For
instance, a more realistic choice of the single-particle potential is the Woods-Saxon
potential, corrected by including an asymmetric term accounting for neutrons excess
in the nucleus,
³ N −Z ´ 1
VW S(p,n) (r ) = −51 ± 33 r −R
. (1.28)
A 1+e a
It is not our scope to discuss further about the Shell Model; more details can be found
in [24–26].
Figure 1.3: The nuclear shell structure; on the left the oscillator shells, on the right the
spin-orbit interaction produces the j -splitting and predict the correct magic numbers.
1.2. DENSITY FUNCTIONAL THEORY 11
1.2 Density Functional Theory

Theoretical many-body approaches to describe nuclear structure have been de-
veloped since the discovery of the neutron. Among those, an important role is played
by nuclear energy density functionals (EDFs). Here, the energy of the system is a func-
tional of the density. The functionals also contain a certain number of parameters,
whose value is fitted to nuclear many-body data, such as binding energies or charge
radii of few nuclei along the chart. There exist three main categories of nuclear energy
density functionals: there are delta-shaped, zero-range (contact) forces, such as the
so-called Skyrme-type interactions [27]; others rely on finite-range, Gaussian-shaped,
forces and go under the name of Gogny-type interactions [28]; finally, relativistic
nuclear energy density functionalss are covariant and yield to a natural inclusion of
the spin degree of freedom to a unique parametrization of time-odd components of
the mean-field [29].
Since the very first works on EDFs, it has been noticed that a density-dependence
of the effective potentials is necessary to the aim of reproducing basic many-body
properties of nuclear structure. Such semi-empirical feature has been explained
in the framework of the nuclear density functional theory (DFT). The charm of the
theory lies in the possibility of predicting, in principle exactly, the ground state
properties of the system under investigation. Moreover, the computational cost that
must be paid is incredibly low. The advantage of treating the density, a function
of three space variables x, y, and z, as a basic variable, in contrast to the involved
many-variable total wave function, is evident. DFT gives then access to the study of
systems that can be very large in the sense of their degrees of freedom. Nowadays,
DFT is the only theory that allows exploring systematically both medium-mass and
heavy nuclei, and it is by far the model that permits to study the highest number of
elements of the nuclear chart (all of them, at least in principle).
The density function is a quantity related to the wave function of the system. Such
quantity is fundamental for the purposes of the thesis and it is the cornerstone of
density functional theory and its inverse problem. The nuclear density is defined as
the number of nucleons per unit of volume in a given state Ψ,
Z Z
n(~
r 1 ) = A · · · |Ψ(~ x A )|2 d σ1 d~
x 1 , . . . ,~ x 2 . . . d~
xA, (1.29)
and it integrates to the total number of nucleons in the system

Z
d~r 1 n(~
r 1 ) = A. (1.30)
A more formal definition of the density function, thought as the local, one-body
restriction of density matrices, is provided in the appendix B. In particular, it can be
shown that such density matrices carry the very same piece of information as the total
wave function, thus justifying the subtitution of basic variable we have mentioned
above.
The Thomas-Fermi (TF) model, as it was proposed in the 1920s, was the first
one to suggest the possibility of determining the energy of an electron system as an
approximate functional of the density, only. The TF energy density functional reads
n(~r) 1 n(~
r 1 )n(~
r2)
Z Z Z Z
5
E T F [n] = C T F d~
r [n(~
r )] 3 − d~r + d~
r 1 d~
r2 . (1.31)
r|
|~ 2 r 1 −~
|~ r2|
For decades, the Thomas-Fermi model has been thought to be oversimplified, be-
cause of its very limited predictive power mainly due to its strong assumptions. The
idea of treating the properties of a system as functionals of the density alone was
later developed by Hohenberg and Kohn; in 1964 [8] they provided three funda-
mental theorems, that showed that the Thomas-Fermi model may be thought as an
approximation of a many-body exact theory: the density functional theory.
1.2.1 Hohenberg-Kohn Theorems

In general, the determination of the ground state energy and of the ground state
wave function of a given fermion system would require both the knowledge of the
number A of particles composing the system, and that of the external potential v(~ r)
to which the system is subject. The first Hohenberg and Kohn theorem justifies the
usage of the local density n(~ r ), alone, as a basic variable, in place of A and v(~r ). It
is trivial that the density uniquely determines the number of particles A. It is much
more remarkable the fact that the external potential v(~ r ) can be identified, up to an
arbitrary constant, by the density n(~ r ). The inter-particle potential, as well as the
kinetic term, are in principle fixed once one considers a given type and number of
fermions.
The First Hohenberg and Kohn Theorem. Given a system of A interacting particles,
described by the Hamiltonian
A ∇2 A
i
X X X
Ĥ = T̂ + V̂ + Ŵ = − + v(~
ri ) + w(~
r i ,~
r j ), (1.32)
i =1 2m i i =1 i 6= j
its A-body wave function Ψ, and the corresponding density (1.29), the non-degenerate
ground state wave function is a unique functional of the ground-state density:
Ψ0 (~ r A ) = Ψ[(n 0 (~
r 1 . . . ,~ r )]. (1.33)
Thereby, the ground state expectation value of each observable is a functional of n(~
r ),
alone.
Proof. The density trivially determines A by quadrature. It is then just necessary to

show, as we have already mentioned, that it determines v(~ r ), too. Suppose there were
two external potential v and v 0 , differing by more than a constant, both related to the
same ground state density n obtained as the solution of the Schrödinger equation.
Then, there would exist two different Hamiltonians, Ĥ and Ĥ 0 , whose ground state
density were the same, even though the normalized wave functions Ψ and Ψ0 would
be different. Let Ψ0 be a trial wave function for the Ĥ problem; then,
E 0 < 〈Ψ0 | Ĥ |Ψ0 〉 = 〈Ψ0 | Ĥ 0 |Ψ0 〉 + 〈Ψ0 | Ĥ − Ĥ 0 |Ψ0 〉

Z
= E 00 + d~ r n(~r )[v(~ r ) − v 0 (~
r )] (1.34)
The same is indeed true if we interchange the dummy primed variables
E 00 < 〈Ψ| Ĥ 0 |Ψ〉 = 〈Ψ| Ĥ |Ψ〉 + 〈Ψ0 | Ĥ 0 − Ĥ |Ψ0 〉

Z
= E 0 + d~ r )[v 0 (~
r n(~ r ) − v(~
r )] (1.35)
By summing the two inequalities, one finds E 0 + E 00 < E 00 + E 0 , which is absurd. There
cannot therefore be two different potentials v providing the same ground state density
n.
1.2. DENSITY FUNCTIONAL THEORY 13
In the same fashion of the Thomas-Fermi model, the total energy can be viewed
as a functional of the density.
The Second Hohenberg and Kohn Theorem. There exists a universal functional
F H K [n], such that the total energy functional reads E v [n] = F H K [n] + V [n] = T [n] +
W [n] + V [n]. The Hohenberg and Kohn functional F H S [n] is universal for a given
particle-particle interaction and for a given number of particles in the system, since it
does not depend on the external potential v(~ r ).
Finally, a third theorem states an energy varational principle for the ground state
energy, involving again the density alone.
The Third Hohenberg and Kohn Theorem. Given some trial density n(~
e r ), such that
n(~
e r)≥0 (1.36)
Z
d~
r n(~
e r)= A (1.37)
The ground state energy is a global minumum of the total energy,
E 0 = E v [n] ≤ E v [n]
e (1.38)
Proof. Because of the first Hohenberg and Kohn theorem, n

e determines its own ve
and its Ψ,
e up to a phase factor. Consider such wave function as a trial for the Ĥ
problem, Z
〈Ψ|
e Ĥ |Ψ〉
e = d~
r n(~
e r )v(~
r ) + F [n]
e = E v [n]
e ≥ E v [n]. (1.39)
If E v [n] is assumed to be differentiable, the third Hohenberg theorem can be

written in the fashion of the method of Lagrange multipliers,
h ³Z ´i
δ E v [n] − µ d~
r n(~
r)− A , (1.40)
where µ is the chemical potential
δE v [n] δF H K [n]
µ= = v(~
r)+ . (1.41)
δn(~r) δn(~ r)
The possibility to have an exact theory urges, and have urged, people to attempt
determining an exact, or properly approximate, structure of the universal functional
F H K [n]. In fact, once we had its explicit form, we could apply density functional
theory to any system. Unfortunately, this reveals to be an incredibly difficult task.
The reason of that is indeed pretty clear, since all of the complications contained in
the many-variable wave function Ψ could not have disappeared at once by moving
to the usage of the simple density function as the basic variable of the many-body
problem. Most of the attempts to provide approximate form of F H K [n] rely on drastic
assumptions, but it is still very remarkable that we have a well-defined procedure for
finding the ground state properties of a possibly huge amount of systems.
1.3 Levy-Lieb Constrained-Search Formulation of DFT

1.3.1 N-Representable and v-Representable Densities
The ground state wave function of a system described by a given Hamiltonian
detemines the ground state density. The Hohenberg and Kohn theorems define then
a one-to-one mapping between the density function and the external potential v(~ r ),
up to an arbitrary constant. If one is able to provide a form of the universal functional
F H K [n], the correct ground state density then determines uniquely the ground state
energy.
We call a density v−representable if it is the related to the antisymmetric ground
state wave function of a Hamiltonian of the type (1.6). It is important then to give a
more precise fomulation of the first Hohenberg and Kohn theorem: There exists a
one-to-one map between the ground state wave function of each many-body quantum
system and the v-representable density of such system. The functionals composing the
energy density functional are at this point defined for v-representable densities, only.
There is a major complication associated to v-representable densities: it is any-
thing but straightforward that a given density is v-representable, since it seems there
does not exist any condition for a trial density to be v-representable1 . It has been
shown that many given densities are not v-representable [31].
Here it comes the relevance of a new formulation of density functional theory,
in terms of densities that satisfy a weaker condition, namely the N -representability
condition. A density is N -representable if it can be obtained from an antisymmetric
wave function. This condition is satisfied by any reasonable one-body local density
and it is a necessary condition for v-representability. By reasonable density we mean
that a density is N -representable if it satisfies the so-called Gilbert conditions:
n(~
r ) ≥ 0, (1.42)
Z
d~r n(~r) = N, (1.43)
Z p
d~r |∇ n(~ r )|2 < ∞. (1.44)
The last condition is a smoothness requirement on the density function form. For
instance, wild oscillations of the density preclude N -representability. In other words,
a N -representable density can be written in terms of N orthonormal orbitals, that
generates n(~ r ) from a single-determinantal wave function [32]. Levy and Lieb inde-
pendently provided such reformulation in 1980, and the next section is devoted to
discuss that.
Before facing the details of the Levy-Lieb formulation of DFT, that implements
N-representability into DFT, let us better comprehend how a N -representable density
can be built in the one-dimensional case x 1 ≤ x ≤ x 2 , starting from N smooth, contin-
uous, and orthonormal orbitals. The constast with the non-constructive definition of
the v-representable densities will be then fully perceptible.
Consider the orbitals
s
n(x) i 2πk Rxx d x n(x)
φk (x) = e 1 N , (1.45)
N
1 Except for some simple dicrete systems [30]

1.3. LEVY-LIEB CONSTRAINED-SEARCH FORMULATION OF DFT 15
where k = 0, ±1, ±2, . . . or k = ± 12 , ± 32 , . . . ; it holds that
n(x)
|φk (x)|2 = , (1.46)
N
and Z x2
d x φ∗k (x)φl (x) = δkl . (1.47)
x1
Any given one-dimensional density can represented as
M
λk |φk |2
X
n(x) = (1.48)
k
with 0 ≤ λk ≤ 1 and M ≥ N . This proves the sufficiency of the Gilbert condition in the
one-dimentional case. Further details on the so-called N -representability problem
can be found in [33].
1.3.2 Levy-Lieb Constrained-Search Formulation

By definition, there exists an infinite number of antisymmetric wave functions
that reproduce the same correct ground-state density; thereby, one could ask how to
distinguish the true ground state wave function Ψ0 from some Ψn0 that also integrates
to the correct ground state density n 0 . In other words, the question is how to point,
among the generic N -representable densities, to the v-representable density that
comes from the true ground state wave function of the Hamiltonian.
This critical theoretical issue of the Hohenberg and Kohn formulation of density
functional theory can be solved by extending the domain of the energy density
functional E v [n] from v-representable densities to the larger set of N -representable
densities. The idea, as suggested by Levy [34] and Lieb, is to exploit the minimum-
energy principle, that uniquely identifies the true ground state wave function
〈Ψn0 | Ĥ |Ψn0 〉 ≥ 〈Ψ0 | Ĥ |Ψ0 〉 = E 0 . (1.49)
The expression above is given by the sum of the expectation value of the universal
functional
F H K [n 0 ] = 〈Ψ0 |T̂ + Ŵ |Ψ0 〉 , (1.50)

= min 〈Ψ|T̂ + Ŵ |Ψ〉 , (1.51)
Ψ→n 0
defined for any v-representable density, and that of the external potential term, cou-
pled to the density. Levy and Lieb formulation writes the minimum energy principle
in order to render more explicit the fact that the variational search is constrained
(compare (1.50) and (1.51)); the trial space for the wave functions is restricted only to
those that reproduce the requested v-representable density n 0 by quadrature. Using
this clever reorganization, it is straighforward to extend the domain of the universal
functional to N -representable densities n
F LL [n] = min 〈Ψ|T̂ + Ŵ |Ψ〉 . (1.52)

Ψ→n
Note that F LL [n 0 ] = F H K [n 0 ].
A double hierarchy minimization for the ground state energy calculation stems
then out,
E 0 = min 〈Ψ|T̂ + Ŵ + V̂ |Ψ〉 (1.53)

Ψ
n h io
= min min 〈Ψ|T̂ + Ŵ + V̂ |Ψ〉 (1.54)
n Ψ→n
n Z o
= min F LL [n] + d~ r n(~
r )v(~
r )] (1.55)
n
The existence of the minimum has been proved by Lieb in 1982 [35]. The Levy-Lieb
constrained-search formulation of DFT removes the original issues associated to
v-representability. The optimization procedure proceeds in two different steps; first
one search the optimal wave function which reproduce a given density. The density
trial space is explored, and at the end of the first minimization one has in its hands a
set of local minima. The sense of the second minimization is simply to identify the
global minimum among those.
1.4. KSDFT 17
1.4 Kohn-Sham Density Functional Theory

The Kohn-Sham scheme gives an insight on the Hohenberg-Kohn theorem, and
especially on the structure of the universal functional F H K [n]. In fact, Kohn and
Sham, by mathematical reorganizing the energy density functional, reformulated
density functional theory [9], turning a theory highly difficult to be handled into a
practical tool for precise calculations.
A set of effective single-particle Kohn-Sham equations are introduced for an auxil-
iary Kohn-Sham system of N non-interacting particles, described by the Hamiltonian
ĤK S = T̂K S + V̂K S . Exploiting the first Hohenberg-Kohn theorem, one asserts that for
any interacting system there exists an unique, local, single-particle potential v K S (~
r)
such that the exact ground state density of the interacting system equals that of a
non-interacting reference system:
N X
r , σ)|2
X
n(~
r ) = n K S (~
r)= |φi (~ (1.56)
i =1 σ
For such reference system, the total ground-state wave function is determinantal
1
ΦK S = p |φ1 φ2 . . . φN |, (1.57)
N!
and the φi can be found as the lowest eigenstates of the single-particle Hamiltonian
HK S :
HK S φi (~
r ) = ²i φi (~
r , σ). (1.58)
The orbitals are unique functionals of the density φi = φi ([n],~ r ), too. The energy
functional reads (let us drop the subscripts for the universal functional)
Z
E [n] = F [n] + d 3 r v K S (~
r )n(~ r) (1.59)
Z
= T [n] + W [n] + d 3 r v K S (~ r )n(~
r) (1.60)
Z
= E K S [n] = TK S [n] + d 3 r v K S (~
r )n(~r) (1.61)
~2 X
Z
=− 〈φi |∇2 |φi 〉 + d 3 r v K S (~
r )n(~
r) (1.62)
2m i
and it is minimized by the correct ground state density of ĤK S . This scheme brings
the universal functional to split into
F [n] = TK S [n] +U [n] + E xc [n], (1.63)
where the first term is the non-interacting kinetic energy, the second term is the local
Hartree term, while the third term is formally defined, throug the previous eqution, as
E xc = T [n] +W [n] −U [n] − TK S [n], and it gathers all the quantum many-body effects.
Thereby, we have a structure for the Kohn-Sham potential
v K S ([n],~ r 0 ) + v xc ([n],~
r ) = v H ([n],~ r) (1.64)
and the exchange-correlation potential is
δE xc
v xc (~
r)= . (1.65)
δn
The Kohn-Sham density functional theory (KSDFT) framework has the advan-
tage of being fully local. Its accuracy depends only on the approximation of the
unknown exchange-correlation energy, which nevertheless is universal. The first level
of approximation, consisting in fully neglecting E xc , delivers the Hartree equations.
Still, the scheme goes far beyond the Hartree mean-field approximation, even only
because it takes into account all the correlation effects and it is, in principle, exact.
Another possible approximation scheme is LDA, in which the exchange-correlation
term is taken equal to that of an infinite uniform system:
Z
LDA
E xc = d~ r n(r )²xc [n]. (1.66)
Among the state-of-art beyond-LDA approximation schemes the most relevant is

the generalized-gradient-approximation (GGA). See [36] for more details. However,
since DFT is an exact non-perturbative theory, it results very difficult to find a clear
direction for systematic improvements of the level accuracy of the calculations.
Chapter 2
Inverse Problem in Nuclear

Kohn-Sham Density Functional
Theory
2.1 Forward and Inverse Problems

The definition itself of inverse problems comes in opposition to that of direct,
or forward, problems. A direct problem deals with the calculation of a quantity, e. g.
the density of some system or its evolution in time, as a function of its causes, for
instance the interaction between the constituent particles or the equation of motion.
Direct problems can usually be expressed through a system of differential equations
that fully determine their evolution. On the other hand, inverse problems typically
arise whenever one tries to perform the indirect observation of some quantity of
interest. Here, the knowledge of some observable is given, and one aims to calculate
the causes that have resulted in that. It is a matter of fact that inverse problems
often present features of non-locality and non-causality. Those features contributes
generating instabilities of the solution of the problem [37].
For instance, consider the inverse heat equation problem, namely the attempt
of estimating an initial temperature distribution, based on the measurement of the
temperature distribution at some final time. It is easy to understand that an infinity
of different initial conditions may have ended up into the same final state: small
changes in the initial temperature could have smeared out in time. On the contrary,
however complex the system could be, the corresponding forward problem is local
and casual: it is guided by the well-known heat equation.
Another example is given by the calculation of the gravitational field of the Earth,
given the knowledge of its density (direct problem) versus the estimate of the density
the Earth given the knowledge of the generated gravitational field (inverse problem).
In the inverse problem case, one must incorporate all available information about
the initial data that one may had prior to the measurement. It is not the scope of the
present thesis to deepen the formalism within which this can be done; more details
can be found, again, in [37].
In the next section we will focus on inverse problems seen as possibly ill-posed
problems and on the regularization techniques one can apply to solve them.
19
20 CHAPTER 2. IKS PROBLEM IN NUCLEAR DFT
2.2 Regularization techniques for Ill-posed Problems

The mathematical definition of well-posed problem was first stated by Jacques
Hadamard. He pointed out three properties that any mathematical model, aiming to
describe a physical system, should respect. They are:
• the existence of a solution of the problem;
• the uniqueness of the solution;
• the continuous dependence of the solution on the input data. This means
that small enough changes in the initial conditions entail small changes in the
solution.1
A problem is well-posed when the above three conditions are fulfilled, while it is said
to be an ill-posed problem if any of those is defective.
Forward problems are in most of cases well-posed ones. Because of that there is
a good chance of being able to implement stable algorithms to produce solutions.
On the other hand, inverse problems are often ill-posed. The existence of an ana-
lytic solution is, in most cases, not guaranteed. It is then necessary to implement
numerical methods to obtain results. When an inverse problem is formulated in
terms of infinite-dimensional function spaces and then discretized for computational
purposes, a discretization error appears. Finite precision easily leads to numerical
instabilities and to unexpected, possibly non-physical, behaviours.
Numerical methods, because of their intrinsic rounding and approximations, can
render inverse problems less ill-posed than they actually are and produce solutions
containing more information than that carried by the input. To ignore the discretiza-
tion errors results in excessively optimistic expectations about the performance of
the method. Those numerical methods that yield to over-optimistic solutions go
under the name of inverse crimes [39]. In those cases, one says that the model is
over-fitting to the input data; that is, a wrong model is reproducing the data too well
in comparison to the knowledge given by the input data.
2.2.1 Tikhonov Regularization

Regularization techniques are developed to get an estimate of the true solution of
an ill-posed problem, as sound and as realistic as possible in relation to the knowl-
edge of the data. Typically, those imply the inclusion in the algorithm of additional
reasonable assumptions, such as the smoothness of the solution.
Let us linger on the details of one specific regularization technique, that will
be later exploited in the definition of our inversion scheme. This is the so-called
Tikhonov regularization [40]. The procedure is indeed a standard one to improve
the solvability of the problem while preventing over-fitting features. The idea of the
regularization method consists in controlling simultaneously both the norm of the
residuals r = Âx − y and the norm of the approximate solution x itself. Let δ > 0 be
a given constant, the so-called regularization parameter. The Tikhonov regularized
solution x δ is the minimum of the functional
F δ (x) = kAx − yk2 + δkxk2 , (2.1)

1 An entire mathematical theory is devoted to the study of catastrophes [38], that is sudden and abrupt
changes in the response of the system to small, smooth changes of the initial conditions.
2.2. REGULARIZATION TECHNIQUES FOR ILL-POSED PROBLEMS 21
provided that such minimum exists. The regularization parameter represents a

Lagrange multiplier and we may think that we are solving the original problem but
constrained by kxk = R, for some R > 0. There exist theorems ensuring that the
Tikhonov regularized solution exists and it is unique. The choice of the value of the
regularization parameter δ is a pivotal issue. One idea is to make use of the Morozov
discrepancy principle [41]. Imagine ² > 0 is an estimate of the norm of the error in the
input vector y; then, any x such that
kAx − yk ≤ ² (2.2)
can be considered acceptable as an approximate solution. If x δ is the regularized

solution, the residual are a function of the regularization parameter, too:
f δ = kAx δ − yk. (2.3)
The Morozov principle states that the parameter δ must be chosen in respect of the
condition f δ = ². That means that the regularized solution should not give residuals
smaller than the noise level of the input data.
The regularization technique we have just presented is strictly valid for linear
problems, only. Nonetheless, the method is sometimes applicable to non-linear
problems along the very same lines we have discussed here (see, again [37]).
2.3 The Inverse Problem In Nuclear Kohn-Sham Den-

sity Functional Theory
2.3.1 Comparing the Inverse and the Forward Problem
In the present thesis, we aim to deduce the form of the Kohn-Sham poten-
tial (1.64), given the knowledge of the density n e of some nuclear system of interest.
The solution of the inverse problem in DFT could reveal to be useful, for instance, to
benchmark the approximate energy functionals available on the market. In this sense,
it can represent an independent strategy to fine tune theoretical models describing
the nuclear structure.
The direct problem in Kohn-Sham DFT, that is the potential-to-density problem,
is widely considered to be well-posed, above all thanks to the Hohenberg and Kohn
theorems and to the Levy-Lieb constrained-search formulation. On the other hand,
although the inverse problem can also be well-posed for some special discretized
systems [42], in most cases errors and lack of detailed information in the input density
lead to the violations of the Hadamard conditions. The density (1.56) determines
a set of single-particle wave functions, up to an overall phase factor; also, unitary
transformations of the Kohn-Sham orbitals wind up in generating the same density,
too. The orbitals in turn identify the form of the Kohn-Sham potential, up to an
arbitrary constant, via the Kohn-Sham equations. Special attention must be paid
then in developing an inversion method that converges to the true solution of the
inverse problem, because non-uniqueness features are a matter of fact.
Let us deepen the detailed differences between the direct and the inverse problem
in Kohn-Sham density functional theory. In general, both of them share the same
set of equations, the Kohn-Sham equations. Still, there are some differences, that
imply the need of developing completely independent algorithms to solve them. Let
us recall the form of the Kohn-Sham equations. For a closed-shell, spin-saturated
system of 2A nucleons,
h ~2 ∇2 i
² j φ j (~
r)= − r ) φ j (~
+ v K S ([n],~ r ), (2.4)
2m
NX
orbs
n(~ r)=2 r )|2 ,
|φ j (~ (2.5)
j =1
v K S ([n],~
r ) = v H ([n],~
r ) + v xc ([n],~
r) (2.6)
where the orbitals are considered to be orthonormal and the external potential is
absent in the nuclear case. In the case of the direct problem, the unknowns are the
set of the wave functions and the density; on the other hand, for the inverse problem
the unknowns are the Kohn-Sham potential and the wave functions. In the forward
problem one must deal with a non-linear eigenvalue problem, since assumptions
for v K S may contain powers of n, ∇n, and so on; instead, in the inverse problem
the eigenvalue problem is fully linear, and non-linearity features are encoded in the
definition of the density. Such subtle difference entails the necessity of using different
methods for the solutions of the two problems. Moreover, the inverse problem is
highly constrained, while the forward problem is free. In fact, in the former case the
knowledge of the density defines several constraints to be respected by the Kohn-
Sham orbitals. Those formal constraints will be made explicit below, in the definition
of our specific inversion algorithm. In both cases, non-linearity requires the choice of
2.3. IKS PROBLEM IN NUCLEAR DFT 23
a smart starting guess of the solution, failing which convergence would not be always
guaranteed.
The inversion problem solution is addressed both to finite and extended systems,
once proper boundary conditions are applied. The main issue here is that in the
framework of the inverse problem we miss a clear piece of information about the
exact boundary conditions that must be applied a priori to the potential. We have
only some phenomenological knowledge of the expected asymptotic behaviour.
However, the density fine details fix somehow strictly the trend of the sum of the
squared orbitals at boundaries. If boundaries are imposed, they should then be in
full agreement with the density constraints. An alternative idea is to let the inversion
method enforce the orbitals’ boundaries to agree with density constraints. That is the
so-called no boundaries density constrained strategy. Such strategy is applied, e.g., in
the van Leeuwen and Baerends (vLB) method; further details on such method can be
found below. Although in those methods no detailed knowledge of the boundaries
is needed, it is quite common to stumble upon inverse crimes. In that case, the
boundary values of the potential may have to be discarded a posteriori. Instead, we
do believe our method could result in a potential more respectful of the real piece
of information carried by the input density. On the other hand, we will make use
of theoretically sound regularization tools, such as the Tikhonov’s, in order to try
reproducing physical results.
In the forward, well-posed, problem, the KS equations can be solved, e. g., self-
consistently, let us say easily to some extent. The errors in the resulting density
will mainly depend on the numerical discretization that will have been used in the
algorithm, on the choice of the convergence condition, and on the particular input
potential. Numerical precision in the inverse problem will also fairly depend on the
discretization of the algorithm, but the quality of the target density represents in
this case a very fundamental limit on the possibility of successfully performing the
inversion. In fact, if the numerical error in a given density is not properly taken into
account, the inversion can produce over-fitting issues or lead to unphysical features
of the output potential. This may be caused by the fact that, in the framework of
the direct problem, the assumed potential is known everywhere at the same level of
detail; in contrast, the density of many system is known in some radial intervals and
extrapolated elsewhere according to some more or less sound theoretical assump-
tions. Actually, results are mainly positive when one tests the algorithm with analytic
formulas of the target density, for which we already know the expected potential, at
least to some extent. On the other hand, when one deals with experimental densities
and treats more-than-one-dimensional systems, much higher attention must be paid.
That is, we believe, our major contribution to the literature, with respect, e. g. to [22]
that deals only with one-dimensional, analytic systems, only.
2.3.2 Inversion methods

Over the last decades many different methods for the solution of the inversion
problem have been proposed [15–22]. Most of them seems to be independent one of
each other, at least to some extent. However, we may think to gather them into two
categories. Some of the methods are based on the optimization of a certain energy
functional, while others consist in a direct algebraic inversion of the Kohn-Sham
equations. In the latter, the potential is recovered within some iterative self-consistent
search of the solution and some density-based quantity is usually exploited to set
the convergence rules. Such quantity also guides the update of the potential in
an iterative process. Instead, in the former the potential is usually defined as the
derivative of the energy functional with respect to the density.
We briefly present now an inversion method that is a fair representative of the
set of the iterative methods, namely the Van Leeuwen and Baerends method (vLB).
The comparison of this method with ours will be very useful in order to deeply com-
prehend the advantages and the drawbacks of both. In vLB mehod, one multiplies
equation (2.4) by φ∗j ,performs a summation over j , divides the result by the density
and finds
2 AX orbs ~2 ∇2
v K S (~
r)= φ∗j (~
r) r ) + ² j |φ j |2 .
φ j (~ (2.7)
n(~
r ) j =1 2m
The formula can be rendered iterative by placing the input density at the denomina-
tor,
n k (~
r) k
v Kk+1 r)=
S (~ v (~r ). (2.8)
e r) KS
n(~
The (k+1)-step density is provided by the solution of the eigenvalue problem, given by
the Kohn-Sham equations relative to the k-step potential. The method is almost self-
explaining: the potential at the (k + 1)-step is increased in regions where n k (~
r ) > n(~
e r)
and vice versa. Normally, according to the usual presentation of the method, the
iteration is successfully terminated
¯ n k (~
r ) ¯¯
max¯1 − ¯ < ², (2.9)
¯
~
r n(~
e r)
where epsilon is some desired threshold. The method has been widely developed and
applied to nuclear systems in [12–14]. Convergence is unreachable unless the starting
guess for the potential is chosen to be very close to the solution. The trial space
explored by the method is therefore quite small. In particular, those works have shown
that the choice of of the convergence criterion is decisive. The criterion suggested
by the literature seems not to be satisfied in most of cases. Further exploration of
possible conditions for stopping the iterations must be then explored. For instance,
in the works mentioned above, the criterion (2.9) is adjusted as
¯ ¯
e r ) − n k (~
max~r ¯n(~ r )¯
¯ ¯
< ², (2.10)
n(~
e r)
to reach convergence. It is not entirely clear why such manual intervention on the
convergence criterion should work better than others. In this method, the initial
guess highly drives the potential behaviour, especially at the boundaries. In fact, the
criterion (2.10) is easily satisfied at large radius, where the values of the density are
infinitesimal. Here the potential profile follows almost perfectly the initial guess. This
could be labelled as an inverse crime, since within the framework of experimental
nuclear densities, we are aware that the values at the boundaries are extrapolated
and quite unreliable. Over-fitting of the potential at large radius is then a real risk. On
the other hand, this feature removes those unphysical behaviours that can appear
when making use of methods that are more sensitive to the input. As we will see in
next section, the vLB method is somehow complementary to the one we have chosen
in the present work of thesis.
Although, from a theoretical point of view, the methods seem to be all well-
defined, we retain that the detailed implementation gives rise to many different
2.3. IKS PROBLEM IN NUCLEAR DFT 25
advantages and drawbacks. We then promote and suggest a combined usage of

different methods, in order to exploit at most simpler methods to complement the
results of the more general and theoretically sound inversion scheme presented right
below.
2.4 Constrained Variational Inversion Method

As presented in [22], the Constrained Variation (CV) method is indeed a promising
inverse method to be applied in Nuclear Physics, especially because it does not
involve the solution of an eigenvalue problem at each iteration. In fact, in many
density-to-potential inversions, such as vLB method, the diagonalization of the
Hamiltonian becomes one of the biggest computational limits whenever the size of
the system increases. Not to solve any eigenvalue problem results in the fact that in
the CV method one obtains wave functions that reproduce the correct target density,
but are eventually an unitary transformation away from the eigenfunctions of the
system. Moreover, with respect to other inversion methods, we believe that this
method is better, since its starting point is not a direct, possibly biased, guess on the
potential form; instead, the starting point for the inversion method is a guess on the
Kohn-Sham orbitals, that are surely closer to the known input, the nuclear density,
and share to some extent a similar functional behaviour. Thus, we avoid any inverse
crime, possibly due to the usage in the inverse problem a piece of information on the
potential coming from the direct problem.
The mathematical tool that is suitable for the implementation of a constrained
optimization procedure is the method of Lagrange multipliers. In the fashion of the
Kohn-Sham scheme, one performs a minimization of the total kinetic energy expec-
tation value with respect to the orbitals, as if the system would be non-interacting:
~2 N
Z
d 3r r )∇2 φ j (~
φ∗j (~
X
f [{φ j }] = − r)
2m j =1
~2 N h
Z i
d 3r ∇ · (φ∗j (~ r )|2
X
=− r )∇φ j (~
r )) − |∇φ j (~
2m j =1
2 Z N
~
d 3r r )|2 ,
X
= |∇φ j (~ (2.11)
2m j =1
where N is the number of filled orbitals in the system. From now on, we will re-
fer to this quantity as the objective function of the optimization. We made use of
an integration by parts to rewrite the kinetic term, so that the dependence on the
wave functions and on their gradients is more explicit. The first term in the second
line gives no contribution if we apply boundary conditions at infinity on the wave
functions.
Two types of constraints c i = 0 are imposed: a density constraint and N (N2+1)
orthonormality constraints.
N
r )|2 − ñ(~
X
c 0 (~
r)= |φ j (~ r ) = 0, (2.12)
j =1
Z
c j k = 〈φ j |φk 〉 − δ j k = d 3 r 0 φ∗j (~
r 0 )φk (~
r 0 ) − δ j k = 0, (2.13)
where j = 1, . . . , N and k = j , . . . , N are an adequate set of quantum numbers. The

orbitals must integrate to the target density and respect orthonormality. Once those
condition are fulfilled, the orbitals can generate a N -representable density (compare
with the discussion in the previous chapter).
There are two equivalent paths we can follow in the development of the method.
First, because the problem we address here is fully time-independent, we can take
2.4. CV INVERSION METHOD 27
the Kohn-Sham orbitals to be real. In the appendix C a formal justification of this

statement is provided. We consider then real wave functions to define the non-
interacting kinetic energy and the constraints, without any loss of generality. We will
show below how, in the practice, this is the same as considering complex orbitals,
since the methods naturally decouples into a real and an imaginary part.
We build then an auxiliary Lagrangian, as prescript by the Lagrange multiplier
method, to turn the constrained minimization into a free one, with respect to the
orbitals’ set and to the Lagrange multipliers. The Lagrange multipliers are formally
defined as
δ f [{φ j }]
λi = . (2.14)
δc i
The core of the inversion lies in the physical meaning of these quantities. The ze-
roth multiplier, related to the constraint (2.12), is directly linked to the Kohn-Sham
potential
δ f [{φ j }] δ
ΦK S ¯ T̂K S ¯ ΦK S
¯ ¯ ®
v K S (~
r)= = (2.15)
δn
PN
δ j =1 |φ j (~
r )| 2
at the minimum.
The multipliers associated to the constraints (2.13) have no plain physical mean-
ing instead; those are the upper triangular part of a symmetric matrix. However, we
could think to diagonalize the matrix containing those multipliers. In this case we
would obtain the sequence of energies ² j of the Kohn-Sham orbitals. The Kohn-Sham
orbitals’ energy have no physical meaning, except for the highest of them, ²max , that
is related to the first ionization energy of the system (DFT-Koopmans’ theorem).
Let us define a cost functional J as the space integral of such auxiliary Lagrangian,
and perform its minimization. Such procedure is totally similar to the fact that the
minimization of the action functional in mechanical system is equivalent to the
solution of the Euler-Lagrange equations
∂L ∂L
−∇· =0 (2.16)
∂φ j (~
r) ∂∇φ j (~
r)
The main advantage is that in such a way we deal with the optimization of a functional
instead of with a multidimensional function.
The cost functional and its associated density, that is the auxiliary Lagrangian,
respectively read
~2 N
Z
r ), ² j k ] = d 3r r )|2 +
X
J [φ j ; v K S (~ |∇φ j (~
2m j =1
Z ³X N ´
+ d 3 r v K S (~
r) r )|2 − ñ(~
|φ j (~ r) +
j =1
Z N X
N ³Z ´
3
²jk d 3 r 0 φ j (~
r 0 )φk (~
r 0) − δ j k ,
X
+ d r (2.17)
j =1 k= j
and
~2 X
N
L [φ j , ∇φ j ; v K S (~
r ), ² j k ] = r )|2 +
|∇φ j (~
2m j =1
N
³X ´
+ v K S (~
r) r )|2 − ñ(~
|φ j (~ r) +
j =1
N X
N ³Z ´
²jk d 3 r 0 φ j (~
r 0 )φk (~
r 0) − δ j k .
X
+ (2.18)
j =1 k= j
The minimum of the kinetic energy, plus the constraints, does coincide by definition
with the minimum of the functional. A necessary condition for the presence of an
extremum is that the Lagrangian satisfies the Euler-Lagrange equations. First, let us
consider the kinetic term:
³ δ δ ´ ~2 X
N h i
−∇· r ))2
(∇φ j (~
δφα δ∇φα 2m j =1
~2
=− ∇2 φα (~
r ). (2.19)
m
The derivation of the constraint terms gives
³ δ δ ´
½ hXN ³ ´ i
−∇· v K S (~
r) φ j (~
r )φ j (~
r ) − ñ(~
r) +
δφα δ∇φα j =1
N XN ³Z ´¾
²jk d 3 r 0 φ j (~r 0 )φk (~r 0) − δ j k
X
+
j =1 k= j
α
hX N i
φ j (~ ²αk φk (~
X
r )φα (~
= 2v K S (~ r)+ r )² j α + r) . (2.20)
j =1 k=α
While deriving the orthogonality term we used the fact that we have only defined the
upper triangular part of the symmetric matrix ² j k ; the sum in this term proceeds as
highlighted in (2.21), with a double counting of the diagonal elements.
²11 ²1α
 
..
.
 
 ↓ 
x2
²αα ²αN 
 

 →  (2.21)
 .. 
 . 
²N N
Equations (2.19) and (2.20) sum up to zero; then, the minimum is identified by the
following set of Euler-Lagrange equations, obtained by artificially introducing φβ (~
r)
and integrating:
~2
Z Z h i
d 3 r φβ (~
r )∇2 φα (~
r)=2 d 3 r φβ (~
r )v K S (~
r )φα (~
r) +
m
Z α
hX N i
+ d 3 r φβ (~ φ j (~ ²αk φk (~
X
r) r )² j α + r) , (2.22)
j =1 k=α
where α = 1, . . . , N and ,β ≥ α. One takes the orbitals coming from the free mini-
mization of the functional, and uses them to get the potential by solving the linear
equations above.
Let us come back to the definition of f [{φ j }] and assume that the orbitals are
fully complex. Then, the matrix ² j k is not symmetric any more, but indeed it is
hermitian. If we refuse to pick the orbitals to be real, a subtle difference stems out in
the derivation of the Euler-Lagrange equations. In fact, we need to calculate both
∂L ∂L
−∇· = 0, (2.23)
∂φ j (~
r) ∂∇φ j (~
r)
∂L ∂L
−∇· = 0. (2.24)
∂φ∗j (~
r) ∂∇φ∗j (~
r)
The result is
~2 N
∇2 φ∗α (~ r )φ∗α (~ ²∗α j φ∗j (~
X
r ) = v K S (~ r)+ r) (2.25)
2m j =α
~2 N
∇2 φα (~ ²αk φk (~
X
r ) = v K S (~
r )φα (~
r)+ r ). (2.26)
2m k=α
The sum of the two equations produces
~2 N
∇2 ℜ[φα (~ ℜ[²αk φk (~
X
r )] = 2v K S (~
r )ℜ[φα (~
r )] + 2 r )].
m k=α
α
hX N i
ℜ[²∗j α φ∗j (~ ℜ[²αk φk (~
X
r )ℜ[φα (~
= 2v K S (~ r )] + r )] + r )]
j =1 k=α
α
hX N i
ℜ[² j α φ j (~ ℜ[²αk φk (~
X
r )ℜ[φα (~
= 2v K S (~ r )] + r )] + r )] , (2.27)
j =1 k=α
that is equation (2.22). On the other hand, the difference of the two equations leads
to the imaginary part of the equations
~2 N
∇2 ℑ[φα (~ ℑ[²αk φk (~
X
r )] = 2v K S (~
r )ℑ[φα (~
r )] + 2 r )]
m k=α+1
hα−1 N i
ℑ[²∗j α φ∗j (~ ℑ[²αk φk (~
X X
r )ℑ[φα (~
= 2v K S (~ r )] + r )] + r )]
j =1 k=α+1
h α−1 N i
ℑ[² j α φ j (~ ℑ[²αk φk (~
X X
r )ℑ[φα (~
= 2v K S (~ r )] + − r )] + r )] (2.28)
j =1 k=α+1
that does not bring along further information if the system is time-independent.
2.4.1 Scaling
In the practice, one may suffer from numerical rounding errors, especially in
those regions where the kinetic energy is small due to the exponential decay of the
wave functions. Therefore, inspired by the density constraint, one rescales the wave
functions as p
φ j (~
r ) = n(~
e r )g j (~
r ). (2.29)
Such redefinition of the orbitals renders the terms in the sum of the kinetic energy all
of the same order of magnitude.
Moreover, as we have already mentioned, there is an intrinsic non-uniqueness
in the problem due to the definition of the density itself. Since the problem is ill-
posed, the competition between orthogonality and density constraints can lead to
non-physical orbitals, in particular when the starting point of the minimization
is not wisely chosen. Typically, one applies a regularization scheme to solve this
issue. We make use of a Tikhonov regularization. Let δ > 0 be a given constant; the
Tikhonov regularized solution is the one that minimizes the functional J plus the
penalty functional
Z N
R[{g j }] = δ d 3 r r )|2 .
X
|∇g j (~ (2.30)
j =1
In the Lagrange picture, the penalty functional can also be read as an extra constraint
on the norm of the gradient of the scaled wave functions k∇g j kL 2 = R, δ being the
corresponding Lagrange multiplier. The orbitals characterized by strong variations
are therefore penalized in the search for the minimum. The value of the penalty
parameter has to be chosen via the Morozov discrepancy principle, which states that
the approximated solution cannot satisfy the constraints more accurately than up to
the numerical noise level of the input density. Thus, the scaled and regularized cost
functional reads
~2 N N
Z p Z
Je[{g j }, v K S , ² j k ] = d 3r r ))|2 + δ d 3 r r )|2 +
X X
|∇( n(~
e r )g j (~ |∇g j (~
2m j =1 j =1
Z N
+ d 3 r v K S (~ r )|2 − 1)
X
r )n(~
e r )( |g j (~
j =1
Z N X
N ³Z ´
d 3r ²jk d 3 r 0 n(~
e r 0 )g j (~
r 0 )g k (~
r 0) − δ j k
X
+
j =1 k= j
2
e r ))2
N h (∇n(~
Z
~ ³ ´
d 3r g 2j (~
X
= r ) + ∇n(~
e r ) · ∇g j (~
r ) g j (~
r)
2m 4n(~
e r)
j =1
i Z N
³X ´
+ (n(~ r )|2 + d 3 r v K S (~
e r ) + δ)|∇g j (~ r )n(~
e r) r )|2 − 1
|g j (~
j =1
Z N X
N ³Z ´
d 3r ²jk d 3 r 0 n(~
e r 0 )g j (~
r 0 )g k (~
r 0) − δ j k .
X
+ (2.31)
j =1 k= j
Again, the solution of the Euler-Lagrange equations
∂L ∂L
−∇· = 0, (2.32)
∂g α ∂∇(g α )
optimizes the functional, and an explicit calculation of those gives
~2 h e r ))2 ∇2 n(~
³ (∇n(~ e r)´ i
g α (~r) − + + ∇n(~
e r ) · ∇g α (~ e r ) + δ)∇2 g α (~
r ) + (n(~ r)
m 4n(~
e r) 2
N
³X α ´
²αk g k (~
X
r )n(~
= 2v K S (~ e r )g α (~
r ) + n(~
e r) r)+ g j (~
r )² j α . (2.33)
k=α j =1
Once more, the artificial introduction of g β (~

r ), plus an integration, provides a matrix
form of those equations,
~2 ³ (∇n(~e r ))2 ∇2 n(~

e r)´
Z h i
d 3 r g β (~
r ) g α (~
r) − + + ∇n(~e r ) · ∇g α (~ e r ) + δ)∇2 g α (~
r ) + (n(~ r)
m 4n(~
e r) 2
Z h N
³X α ´i
= d 3 r g β (~ ²αk g k (~
X
r )n(~
e r ) 2v K S (~
r )g α (~
r)+ r)+ g j (~
r )² j α , (2.34)
k=α j =1
where α = 1, . . . , N , β ≥ α.
2.5 Insertion of Spherical Symmetry into the Constrained

Variational Method
2.5.1 The Spherical Hamiltonian
In order to lay the ground for solving the inverse Kohn-Sham problem relative
to nuclear densities, let us consider systems characterized by spherical symmetry.
This means the Hamiltonian will be invariant under rotations, and the Kohn-Sham
potential will be central.
The assumption of spherical symmetry implies that the Hamiltonian commutes
with the squared angular momentum and with one of its components, say the projec-
tion along the z-axis,
[ Ĥ , L̂ 2 ] = [ Ĥ , L̂ z ] = 0. (2.35)
Therefore, the Hamiltonian separates as
p̂ r2L̂ 2
Ĥr = + + V̂ (r ) (2.36)
2 2r 2
1 ³ ∂2 2 ∂ L2 ´
=− + − + V̂ (r ). (2.37)
2 ∂r 2 r ∂r r 2
Since we are trying to generalize our inversion model as much as possible, let us
consider systems characterized by a spin degree of freedom, too. We require the
Hamiltonian to commute with the square of the spin operator S, together with one of
its components
[ Ĥ , Ŝ 2 ] = [ Ĥ , Ŝ z ] = 0. (2.38)
In such a way we have chosen the so-called |nl m l sm s 〉 representation, namely the
decoupled basis. The components of the wave functions on the decoupled basis, in
the position space, read then
u nl (r )
ϕnl ml sm s (~
r , σ) = Yl ml (θ, φ)χsm s (σ). (2.39)
r
The functions Yl ml (θ, φ) are the well known spherical harmonics, which form a
complete orthonormal set on the sphere S2 . They are defined by
s
(2l + 1) (l − m l )! ml
Yl ml (θ, φ) = (−1)m P (cos θ)e i ml φ , (2.40)
4π (l + m l )! l
m
where P l l (x) are the generalized Legendre polynomials, l = 0, 1, . . . and m l = −l , . . . , l .
The requirement on the wave functions to be eigenfunctions of the angular momen-
tum, characterized by definite values of the quantum numbers l and m l , completely
fixes their angular dependency.
In the following, it will be useful to write the orbitals in the coupled basis |nl s j m j 〉,
the operator J = L + S standing for the total angular momentum. Such alternative set
of quantum numbers is more convenient than the other, because an overall rotational
invariance (orbital momentum and spin) implies that that j and m j are constants of
motion. Moreover, the great importance of the spin-orbit coupling
Jˆ2 − L̂ 2 − Ŝ 2
H 0 (r ) = −α(r )L̂ · Ŝ = −α(r ) , (2.41)
2
2.5. CV METHOD IN SPHERICAL SYMMETRY 33
in nuclear systems guides our basis choice; the spin-orbit term is diagonal in the
coupled representation.
We decide then to simultaneously diagonalize the angular momentum and the
spin operator, together with the total angular momentum and one of its components
[ Ĥ , L̂ 2 ] = [ Ĥ , Ŝ 2 ] = [ Ĥ , Jˆ2 ] = [ Ĥ , Jˆz ] = 0. (2.42)
The orbitals on this basis read
u nl j (r )
ϕnl s j m j (~
r , σ) = [Yl (θ, φ) ⊗ χs (σ)] j m j . (2.43)
r
Either description, decoupled and coupled, is complete. Since the two bases are
fully equivalent, they must be connected by a unitary transformation:
X
|nl s j m j 〉 = 〈l m l sm s |l s j m j 〉 |nl m l sm s 〉 . (2.44)
ml m s
The matrix bridging the two representations is composed by the Clebsch-Gordan

coefficients; those coefficients are independent of the quantum number n, since the
transformation regards the angular parts of the wave function, alone. In the following,
we will make use of some of their properties:
〈l s j 0 m 0j |l m l sm s 〉 〈l m l sm s |l s j m j 〉 = δ j j 0 δm j m 0
X
(2.45)
j
ml m s
〈l m l0 sm s0 |l s j m j 〉 〈l s j m j |l m l sm s 〉 = δml m 0 δm s m s0 .
X
(2.46)
l
jmj
A detailed descussion about how the Clebsch-Gordan coefficients properties and

how they can be explicitely caluclated can be found, e.g., in [24]. The Clebsch-Gordan
coefficients are computed up to an overall phase; the standard convention is to pick
them up as real and satisfying the following symmetry properties
〈l m l sm s |l s j m j 〉 = (−1)l +s− j 〈l − m l s − m s |l s j m j 〉 , (2.47)

〈l m l sm s |l s j m j 〉 = (−1)l +s− j 〈sm s l m l |l s j m j 〉 , (2.48)
s
s−m s 2j +1
〈l m l sm s |l s j m j 〉 = (−1) 〈s − m s j m j |s j l m l 〉 . (2.49)
2l + 1
The insertion of an identity resolution into the coupled decomposition allows to re-
cover the well known structure, easy indeed to be treated, of the spherical harmonics.
The cost to be paid in order to do that is the computation of the Clebsch-Gordan
coefficients. Luckily, we will show that we can get rid of the CG coefficient through
some manipulations and exploiting the symmetries we have assumed.
u nl j (r )
ϕnl s j m j (~
r , σ) = [Yl (θ, φ) ⊗ χs (σ)] j m j (2.50)
r
u nl j (r ) X
= 〈θφσ|l m l sm s 〉 〈l m l sm s |l s j m j 〉 (2.51)
r ml m s
u nl j (r ) X
= 〈l m l sm s |l s j m j 〉 Yl ml (θ, φ)χsm s (σ). (2.52)
r ml m s
2.5.2 The Spherical Form of the Cost Functional

We are ready to define the quantities of our inversion methods in terms of spheri-
cal coordinates. First, let us calculate the nuclear density:
(q)
n (q) (~ r , σ)|2
X
r)= |ϕnl s j m (~
j
nl s j m j
2(q)
X u nl j (r )
〈l s j m j |l m l0 sm s0 〉 〈l m l sm s |l s j m j 〉
X
=
nl j m j r2 m l m s m l0 m s0
Yl∗m 0 (θ, φ)Yl ml (θ, φ) χ∗sm 0 (σ)χsm s (σ)]

X
l s
s
2(q)
u nl j (r ) 2j +1
〈l sl m l0 | j m j s(−m s0 )〉 〈 j m j s(−m s )|l sl m l 〉
X X
=
nl j r2 m l m s m l0 m s0 m j
2l + 1
[Yl∗m 0 (θ, φ)Yl ml (θ, φ)δm s m s0 ]

l
2(q)
X u nl j (r ) X 2 j + 1
= 2
δml m 0 Yl∗m 0 (θ, φ)Yl ml (θ, φ)
nl j r m m 0 2l + 1 l l
l l
2(q)
u nl j (r ) 2j +1 X ∗
Y (θ, φ)Yl ml (θ, φ)
X
=
nl j r2 2l + 1 ml l ml
2(q)
X 2 j + 1 u nl j (r )
= . (2.53)
nl j 4π r2
In the second step we have used the completeness of the spinors; moreover, we
have exploited the symmetry properties of the Clebsch-Gordan coefficients in order
to make the m j -degeneracy explicit. Finally, in the last step we have exploited a
resolution of the identity and the following property of the spherical harmonics:
l 2l + 1 0
Yl∗m (θ1 , φ1 )Yl m (θ2 , φ2 ) = P (cos ω),
X
(2.54)
m l =−l 4π l
cos ω = cos θ1 cos θ2 + sin θ1 sin θ2 cos(φ1 − φ2 ), (2.55)
where cos ω = 1 for θ1 = θ2 , φ1 = φ2 , and P l0 (1) = 1.

The calculation of the kinetic term passes through the transformation of the
Laplace operator in spherical coordinates,
∂2 f 2 ∂ f Λ2 f
∇2 f = + + 2 (2.56)
∂r 2 r ∂r r
1 ∂2 Λ2 f
= (r f ) + 2 , (2.57)
r ∂r 2 r
whose angular dependence is entirely contained in the Legendre operator,
1 ∂ ³ ∂ f ´ ∂2 f
Λ2 f = sin θ + . (2.58)
sin θ ∂θ ∂θ ∂φ2
The eigenfunctions of the Legendre operator are exactly the spherical harmonics
Yl m (θ, φ), with eigenvalues −l (l + 1). By comparing equations (2.37) and (2.56), it is
straightforward to notice that, from a physical point of view, the Legendre operator
represents minus the angular momentum operator in the position space. The de-
composition and the radial rescaling that we have applied to the wave functions are
explicitly thought to simplify the way the Laplace operator acts on them. The total
kinetic energy expectation value, that the first term of the cost functional, transforms
as
X³ ~2 ´ Z 2 X X ∗(q) ³ 1 ∂2 Λ2 ´ (q)
JI = − r drdΩ ϕ 1 σ)
r ,~
(~ r + ϕ r , σ)
(~
q 2m q σ nl j m j nl 2 j m j r ∂r 2 r 2 nl 12 j m j
(q)
X³ ~2 ´ Z 2 X X u nl j (r ) X 1
= − r drdΩ 〈l m l0 m s0 | j m j 〉 Yl∗m 0 (θ, φ)χ∗1 0 (σ)
q 2m q σ nl j m j r 2 l 2 ms
m0 m0 l s
(q)
³ 1 ∂2 Λ2 ´ u nl j (r ) 1
〈l m l m s | j m j 〉 Yl ml (θ, φ)χ 1 m s (σ)
X
r+
r ∂r 2 r2 r ml m s 2 2
(q) 2 (q)
X³ ~2 ´ Z 2 X X³ u nl j (r ) ∂ u nl j (r ) 2(q) l (l + 1) ´
= − r dr − u nl j (r )
q 2m q nl j m j r 2 ∂r 2 r4
1 1
Z
〈l m l m s | j m j 〉 〈 j m j |l m l0 m s0 〉 d ΩYl∗m 0 (θ, φ)Yl ml (θ, φ)
X
m l m s m l0 m s0
2 2 l
X ∗
χ 1 0 (σ)χ 1 m s (σ)
σ 2 ms 2
X³ ~2 ´ Z X ³
(q) ∂2 (q) 2(q) l (l + 1) ´
= − d r (2 j + 1) u nl j (r ) 2 u nl j (r ) − u nl j (r )
q 2m q nl j ∂r r2
1 1
〈l m l m s | j m j 〉 〈 j m j |l m l0 m s0 〉 δml m 0 δm s ,m s0
X
ml m s m0 m0
2 2 l
l s
X ~2 Z ³ ∂
(q) 2(q) l (l + 1) ´
d r (2 j + 1) | u nl j (r )|2 + u nl j (r )
X
= . (2.59)
q 2m q nl j ∂r r2
The key idea is that once the Laplace operator acts on the wave functions, we can
exploit the orthogonality of the spinors, of the spherical harmonics, and that of the
Clebsch-Gordan coefficients. We would like to remark that the coordinate transfor-
mation of the kinetic term gives rise to two contributions; the first is the proper radial
kinetic term, while the second represent a centrifugal fictitious potential.
The potential term, that is the second term in the functional, reads
XZ (q)
h X
(q)
i
r 2d r d Ω r , σ)|2 − n
e(q) (r )
X
J II = v K S (r ) |ϕ 1 (~ (2.60)
q σ nl j m j
2
nl j m j
XZ (q)
hX
2(q)
i
= d r v K S (r ) (2 j + 1)u n j l (r ) − 4πr 2 n
e(q) (r ) . (2.61)
q nl j
Here we made use of the nuclear density expression (2.53). It is quite immediate to
read out the density constraint from this term,
2(q) !
(2 j + 1)u n j l (r ) − 4πr 2 n
e(q) (r ) = 0.
X
(2.62)
nl j
Finally, the orthonormality term is given by
XZ hZ
ϕ∗α0 (~
r 0 , σ0 )ϕα (~
i
(q) 2
r 2d r d Ω ²αα0 r 0 d r 0 d Ω0 r 0 , σ0 ) − δαα0
XX X
J III =
q σ αα0 σ0
XZ (q)
r 2d r d Ω ²(nl j m
X X X
= 0 l 0 j 0 m0 )
j ),(n
q σ nl j m j n 0 l 0 j 0 m 0 j
j
(q) (q)
hZ u nl j (r 0 )u n 0 l 0 j 0 (r 0 ) 1 1
0 02
〈l m l m s | j m j 〉 〈 j 0 m 0j |l 0 m l0 m s0 〉
X
dr r
r 02 ml m s m0 m0
2
s
2
l
Z i
d Ω0 Yl∗0 m 0 (θ 0 , φ0 )Yl ml (θ 0 , φ0 ) χ∗1 (σ0 )χ 1 m s (σ0 ) − δnn 0 δl l 0 δ j j 0 δm j m 0
X
0
l
σ0 2 ms 2 j
XZ (q)
r 2d r d Ω ²(nl j m
X X X
= 0 l 0 j 0 m0 )
j ),(n
q σ nl j m j n 0 l 0 j 0 m 0 j
j
hZ
(q) (q) 1 1 1
d r 0 u nl j (r 0 )u n 0 l 0 j 0 (r 0 ) 〈l m l m s |l j m j 〉 〈 j 0 m 0j |l 0 m l0 m s0 〉
X
m l m s m l0 m s0
2 2 2
i
δl l 0 δml m 0 δm s m s0 − δnn 0 δl l 0 δ j j 0 δm j m 0
l j
XZ 2 X X X (q)
= r drdΩ ²(n j m ),(n 0 j 0 m 0 )
j
q σ nl j m j n 0 j 0 m 0 j
j
hZ
(q) (q) 1 1 i
d r 0 u nl j (r 0 )u n 0 l j 0 (r 0 ) 〈l m l m s | j m j 〉 〈 j 0 m 0j |l m l m s 〉 − δnn 0 δ j j 0 δm j m 0
X
ml m s 2 2 j
XZ X (q)
r 2d r d Ω ²(nl j m ),(n 0 l j 0 m 0 ) δ j j 0 δm j m 0
X X
=
j j
q σ nl j m j n 0 j 0 m 0 j
j
hZ i
(q) (q)
d r 0 u nl j (r 0 )u n 0 l j 0 (r 0 ) − δnn 0
XZ X (q) X hZ
(q) (q)
i
= 4πr 2 d r ²nn 0 (2 j + 1) d r 0 u nl j (r 0 )u n 0 l j (r 0 ) − δnn 0 . (2.63)
q nn 0 lj
In other words, the physical request of orthogonality of the wave functions involves
just their radial part. In fact, the orthogonality of the angular and spin part is implicitly
hidden in the decomposition we have made. Note that as well as in the general
(q)
calculation, the matrix ²nn 0 is symmetric, and we can pick the n 0 ≤ n case, only. The
orthogonality constraints are therefore
hZ i
(q) (q) !
d r 0 u nl j (r 0 )u n 0 l j (r 0 ) − δnn 0 = 0
X
(2 j + 1) (2.64)
lj
Thus, summing up these three terms we get the cost functional in spherical
symmetry
(q) (q)
J [{u nl j (r )}, v K S (r ), ²nn 0 ] = J I + J II + J III
Xn ~2 Z ³ ∂
(q) 2(q) l (l + 1) ´
d r (2 j + 1) | u nl j (r )|2 + u nl j (r )
X
=
q 2m q nl j ∂r r2
Z hX i
(q) 2(q)
+ d r v K S (r ) (2 j + 1)u n j l (r ) − 4πr 2 n e(q) (r )
nl j
Z hZ io
(q) (q) (q)
4πr 2 d r d r 0 u nl j (r 0 )u n 0 l j (r 0 ) − δnn 0
XX
+ (2 j + 1)²nn 0 . (2.65)
nl j n 0
The Euler-Lagrange equations produce the linear system whose solution mini-
mize the cost functional and satisfies the imposed constraints. We apply the spherical
symmetry to the linear system,
(q)
Z
~2 h ∂2 u nl j (r ) l (l + 1) i
(q) (q)
d r u n 0 j l (r ) − u nl j (r )
mq ∂r 2 r2
Z
(q) (q) (q)
=2 d r u n 0 j l (r )v K S u nl j (r )
Z n
hX NX
max i
(q) (q) (q) (q) (q)
+ d r u n 0 j l (r ) ²ns u sl j (r ) + ²t n u t l j (r ) (2.66)
s=1 t =n
The rescaling of the wave functions and the insertion of a penalty functional are trivial
and follow the very same procedure we have already done in the general calculation.
In particular, the rescaling we have chosen is inspired by the form of the density
constraint (2.62): p
u nl j (r ) = 4πr 2 ng
e nl j (r ). (2.67)
Its substitution into the cost functional reads
(q) (q)
Je[{g nl j (r )}, v K S (r ), ²nn 0 ] =
Xn ~2 Z (q)
h³ ∂n(r
e ) ∂2 n(r
e ) r 2 ³ ∂n(r
e ) ´2 ´ (q)
+r2
X
d r (2 j + 1)4πg nl j (r ) r − g nl j (r )
q 2m q nl j ∂r ∂r 2 e ) ∂r
4n(r
(q) (q)
e ) ´ ∂g nl j (r ) ∂2 g nl j (r ) l (l + 1)
2 ∂n(r
³ ´
2 (q)
+ 2r n(r
e )+r e ) + δ)
+ (r n(r − n(r )g (r )
∂r ∂r ∂r 2 r2 nl j
e
Z hX i
(q) (q)
+ d r v K S (r ) (2 j + 1)g nl j (r ) − 1
nl j
Z hZ io
(q) (q) (q)
4πr 2 d r d r 0 4πr 2 n(r
e 0 )g nl j (r 0 )g n 0 l j (r 0 ) − δnn 0
XX
+ (2 j + 1)²nn 0 . (2.68)
nl j n 0
2.5.3 Spherical Bondary Conditions

The reduction of the original three-dimensional problem to a unidimensional
one, via the change to spherical coordinates, entails the appearance of extra boundary
conditions that the solution must respect.
First, the wave functions have to be normalizable, and this imply that its integral
has to converge in the origin. This is possible if
u(r → 0) ∼ r −β , (2.69)
with β < 1/2. However, one can prove that, if the potentials to which the system is
subject to are not extremely pathological, the condition above gets more restrictive,
that is
u(r → 0) = 0. (2.70)
With such a piece of information, a more detailed study of the behaviour of the
wave functions in the origin can be done. As r → 0, the centrifugal term is always
dominating the Hamiltonian if we require the energy spectrum to be lower bounded.
The Schrödinger equation close to the origin then behaves as
d 2 u(r ) l (l + 1)
− u(r ) = 0. (2.71)
dr 2 r2
The equation is solved by polynomials of the type
u(r ) = Ar l +1 + Br −l . (2.72)
It is clear that if we apply the integrability conditions we have just discussed above,
the parameter B = 0, and u(r → 0) ∼ r l +1 .
At infinity, the behaviour of the wave functions depends on that of the potential.
Indeed, if V (r → ∞) → 0, the asymptotic of the wave functions reflects that of a free
particle. In fact, in this case the Schrödinger equation at infinity becomes
d 2 u(r )
= −2E u(r ). (2.73)
dr 2
Bounded states (E < 0) then behave as
p
2|E |r
u(r → ∞) ∼ e − . (2.74)
Last but not least, note that in contrast with the unidimensional case, where the
existence of at least one bound state is always guaranteed, this is not the case for three-
dimensional systems. In fact, it is known that the ground state of a system subject
to an one-dimensional system exists, is even, and the excited states alternate their
parity. Nonetheless, even though we have seen that the Hamiltonian of a spherical
symmetric system can be treated as one-dimensional (keeping into account the
restriction r > 0), the boundary condition (2.72) implies that only odd solutions
are acceptable, and a ground state may not exist. Finally, the repulsiveness of the
centrifugal potential certainly involves that the number of possibly bound states
decreases as l rises.
2.6. PARAMETRIZATIONS OF EXPERIMENTAL NUCLEAR DENSITIES 39
2.6 Parametrizations of Experimental Nuclear Densities

With the definition of a theoretically sound density-to-potential method at hand,
we are left with the problem of how to obtain the input density for the inversion
scheme.
Indeed, one can test the robustness of the model on analytic formulas of the
density, as obtained by solving the Schrödinger equation for a given potential and
inter-particle interaction. This is what we will discuss about in the first part of the
next chapter of the thesis.
It is much more involving and interesting to use some realistic densities as inputs.
Those can be obtained in seveveral ways. The second part of next chapter will be
devoted to the analysis of the results of the derivation of a Kohn-Sham potential for
realistic calculations or experimental data in magic, spherical, nuclei.
We must remark that it is not an easy task to get experimental data on nuclear
densities; the literature is poor for which regards neutron densities, while proton den-
sities must be obtained from charge proton density. Nuclear experimental densities
are usually deduced by measuring the electron or proton elastic scattering by nuclei.
The main reference that collects the experimental proton charge densities of nuclei
is a paper by De Vries et al. [1]. Electron elastic scattering is the most easy-access
method to study nuclear structure, thanks to the fact that the form of the electric
force is known. The link between the differential cross section and the form factor
of the target depends on the nature of the forces acting between the probe and the
target. In Plane-Wave Born Approximation, valid for small atomic number Z , for the
case of Coulomb forces, the relation is given by
d σ(q)
= σMott |F (q)|2 , (2.75)
dΩ
where σMott is the Mott cross section [43], that is typical of the scattering of an
electron beam from a target with the size of a nucleus. In the case of high Z , a more
involved model (Distorted-Wave Born Approximation) must be used, and details of
the proportion between the two quantities can be found in [44]. Then, the form factor
of the nucleus depends on the charge density as
Z ∞ sin(qr )
F̂ (q) = 4π r dr n charge (r ), (2.76)
0 q
The charge density can be obtained as the inverse Fourier transformation of the
above equation. The energy scale of the electron-proton interaction is low enough to
allow stopping perturbative calculations of cross sections at the first order. On the
other hand, there are at least two evident drawbacks in such type of measurements
of the density:
• electrons interact with protons only, not with neutrons. Therefore, via elec-
tronic scattering experiments, it is possible to get proton charge densities, only.
A method to calculate the proton barionic density from the proton charge
density is therefore needed. As it is shown in appendix D, since the proton form
factor is known, this can be done through a deconvolution.
• the experimental analysis of the nuclear form factor is restricted to a finite

interval of momentum transfers. It is thereby possible to get sound information
only relatively to the exterior part of nuclei. The interior, as well as the long tail,
part of the densities can only be extrapolated through a fit, usually exploiting
a minimum χ2 method and an assumption of the form of the density. Those
regions reveal then to be characterized by high uncertainties, and they are
rather unreliable.
Zenihiro et al. [2] managed to provide the neutron density of few isotopes of lead
and tin, through the analysis of elastic proton scattering. Due to the fact that the
knowledge of the strong interactions in the mean is all but complete, differential
cross sections measured by hadron scattering are in general characterized by much
higher and systematic uncertainties than the correspondent electron cross sections.
Here, it is necessary to rely on an effective interaction, in order to provide the relation
between the form factor and the differential cross section. The energy of the proton
beam in those experiments is around 300 MeV. This is a good value both to avoid the
production of heavy mesons in the scattering reactions and to extract information
about the shape of the surface of the nucleus. Again, the density IS extrapolated in
the other regions.
Extrapolations are obtained by assuming a parametrization of the density, that
allows for a fit to nuclear charge radii.
A possible choice is to fit the density to a two- or three- parameters Fermi (2pF,
3pF) function:
1
n (2pF ) (r ) = n 0 r −c , (2.77)
1+e z
2 2
(3pF )
1 + wc 2r
n (r ) = n 0 , (2.78)
r 2 −c 2
1+e z2
where c, z, and w are fit parameters. The main advantage of those kinds of fits is that
the tails of the densities result more similar to those theoretically expected, and fall
to zero with the correct rough behaviour. However, features of model-dependence
are often present: results may highly differ according to the choice of the parameters.
At present, the majority of results are analysed are Fourier-Bessel (FB) analysis
and sum of Gaussians (nSoG). Details on the possible choices of the parametrizations
of nuclear densities, with their advantages and drawbacks, are available in [45].
In the present thesis we made use of the nSoG parametrizations, only. In fact FB
analysis in not adequate for our purposes, since it present a cutoff to zero at some
radial value. The discontinuity of the input would then preclude the usage of our
method, for which quanties of interest must be at least twice continuous.
It must be highlighted that in the experimental literature, the term model in-
dependent is used to address those parametrization that are not highly dependent
from the specific value of the parameters. For instance, if the number of Gaussian
used in nSoG is large enough, results are independent from that. On the contrary,
a given choice of the parametrization entails a specific, different,
p extrapolation of
the densities in the outer regions. In nSoG, the width γ/ 2 of the Gaussians is set
to be equal to the smallest width of the radial wave functions, as obtained through
Hartree-Fock calculations. The values of n(r ) at different values are decoupled; we
mean that the rapid decrease of the Gaussians allows to assume that the density at
2.6. PARAMETRIZATIONS OF EXPERIMENTAL NUCLEAR DENSITIES 41
some point is parametrized, to some extent, by the closest Gaussian, only.
n A (p,n) n (r −R i )2 (r +R i )2
− −
o
(nSoG) i
X
n (p.n) (r ) = e γ2 +e γ2 (2.79)
i =1 γ3
p Z eQ i
Ai = (2.80)
3 R2
2π (1 + 2 γ2i )
2
NQ i
A ni = (2.81)
3 R2
2π (1 + 2 γ2i )
2
(2.82)
For which regards De Vries’ densities, this represents the parametrization of the
proton charge density; it is necessary to apply the deconvolution we have mentioned
above in order to obtain the proton density,
2 2
n A n³ r − R R i ´ − (r −R i) r + R i R i ´ − (r +R i) o
i i
n p(SoG) (r ) =
X
+ e β2 + ( − e β2 . (2.83)
i =1 eβr β2 γ2 β2 γ2
Instead, in Zenihiro’s article, parametrization (2.80) with (2.81) is obviously di-

rectly referred to neutron densities. The parameters R i represent the centres of the
Gaussian, while Q i are related to the fraction of charge (or of particles) brought by
each Gaussian. Indeed, ni=1 Q i = 1.
P
The main drawback of such parametrization is that it assumes a completely

wrong analytical shape for the tails of the densities. We will see how, within the nSoG
parametrization, the extrapolated tail of the density will produce a potential that is
known to produce Gassians wave function: the harmonic oscillator potential.
Chapter 3
Implementation of the
Constrained Variational Method
and Inversion Results
The chapter is devoted to present the results of the CV method for the solution of
the inverse Kohn-Sham problem. The first section discusses the numerical imple-
mentation of the method. Then, a series of tests of increasing complexity is presented.
Finally, we analyse and discuss the results of the inversion when the input density is
a nuclear one.
3.1 Implementation of the Constrained Variational Method

The perspective we have followed during almost one year of developments has
been to provide an algorithm applicable to inverse problems that are as general as
possible. In the end, we managed to build a software that can be used for solving three-
dimensional spherical problems. However, the formalism of the method, as well as
its implementation, are explicitely thought in order to make further generalizations
natural.
We make use of the IPOPT (Interior POint OPTimizer) library1 , projected to solve
large-scale non-linear optimization problems of the kind
min f (~
x) (3.1)
~
x ∈Rn
s.t. g iL ≤ g i (~
x ) ≤ g iU (3.2)
x Lj ≤ x j ≤ xU
j , (3.3)
where x j are the variables of the problem, f is the objective function to be optimized,
and g i are constraints. The functions can be linear or non-linear, convex or non-
convex, but should be at least twice continuously differentiable. In general, the
constraints have lower and upper bounds, but those can collapse one on the other to
specify equality constraints. The software implements an interior-point filter line-
search algorithm [46], that aims to find a local solution of the optimization problem.
1 https://projects.coin-or.org/Ipopt
43
44 CHAPTER 3. IMPLEMENTATION AND INVERSION RESULTS
We make use of the software to perform the numerical constrained-minimization

of the objective function. In our case, this would be (2.11) for a non-scaled, non-
regularized cartesian problem, the first term of (2.31) for a scaled and regularized
cartesian problem, and the first term in (2.68) for a spherical problem. The program
requires, as a further input with respect to the target density, a starting guess for the
set of the Kohn-Sham orbitals. In the section of the results, we will discuss, case by
case, which has been the particular choice of the starting point for the optimization.
The software begins modifying such initial guess for the orbitals, aiming to make new
trials minimize the objective function and to respect the constraints of the problem.
The routine computes then, at each step, the objective function and its gradient, that
is used for the choice of the search direction. The program also calculates at each
step the constraints (2.12) and (2.13), their violation with respect to zero and their
gradient. Second derivatives of the quantities listed above are passed to the software
by building the Hessian of the Lagrangian of the problem, that is
x ) X ∂2 g i (~
∂2 f (~ x)
H kl [ f (x), g i (x); λi ] = + λi , (3.4)
∂x k ∂x l i ∂x k ∂x l
where λi are the Lagrange multipliers correspondent to the constraints g i . Finally,

the search must be restricted to a compact space, to guarantee the existence of an
extremum and to satisfy the hypotheses of the Lagrange Multiplier Method. We will
discuss the values of the intervals defining such compact space, case by case, in the
following.
In general, if the new trial for the optimization improves the objective function
value, or if it improves the violation of the constraints, the new guess is accepted; else,
the iteration is restored to the previous step. The process is iterated up to when a
local minimum is found, that is when small variations of the orbitals do not produce
any decrease of the objective function, within a user-chosen tolerance. With such
piece of information, the linear system of Euler-Lagrange equations, given by (2.22),
can be solved.
As we have already mentioned, a discretization of the variables of the problem
is required for the numerical implementation of the problem. Therefore, the space
is divided into a grid of equidistant points. The variables of the problem, that is the
Kohn-Sham orbitals, are defined, point by point, on the spatial grid. The choice of
the methods for deriving and integrating becomes then crucial for the efficiency of
the minimization.
More details on the implementation of numerical derivatives, integrals and on
the implementation of the quantities mentioned above, are available in appendix E.
3.2. UNIDIMENSIONAL NUMERICAL TESTS 45
3.2 Unidimensional Numerical Tests

3.2.1 Harmonic Oscillator Density
As a first test, we apply our inversion scheme to a unidimensional system subject
to an harmonic external field, that is
1
v ext (x) = mω2 x 2 , (3.5)
2
For such a simple system we have an analytic expression for all of the eigenfunctions,
³r mω ´ mωx 2
φn (x) = c n Hn x e − 2~ (3.6)
~
r
1 ³ mω ´ 41
cn = . (3.7)
2n n! π~
The Hermite polynomials are formally defined as
2 d n −x 2
Hn (x) = (−1)n e x e , (3.8)
d xn
and can be implemented according to the recurrence relation
H0 (x) = 1,
H1 (x) = 2x,
Hn (x) = 2x Hn−1 (x) − 2(n − 1)Hn−2 (x) (3.9)
for n ≥ 2. Therefore, depending on the number of filled orbitals, we get an analytic

expression of the target density by using equation (1.56).
One-orbital Inversion
We begin by considering a one-orbital system. The target density reads

³ mω ´ 1 mωx 2
2 −
e = φ20 (x) =
n e ~ . (3.10)
π~
Thus, the system we must solve is
(
ϕ21 (x) − n(x)
e =0
~2 R 2 2
(3.11)
m d x ϕ1 (x)∇ ϕ1 (x) = 2 d x (v K S (x) + ²11 )ϕ1 (x)
R
Since just one orbital is occupied, the requirement of normalization p is hidden in

the first equation. Also the only orbital is determined by ϕ1 (x) = ± n(x). e This
happens every time we consider one-orbital systems. Such feature renders one-
orbital inversions somehow special: an analytical solution of the inversion is always
possible and the inversion problem is actually well-posed (in particular, the solution
is unique) if we discard the negative solution and forget the arbitrary constant of the
potential. With the Kohn-Sham orbital at hand, one computes its Laplacian,
d 2 h³ mω ´ 14 − mωx 2 i
∇2 ϕ1 (x) = e 2~
d x 2 π~
³ mω ´h³r mω ´2 i
= ϕ1 (x) x −1 ,
~ ~
and gets the Kohn-Sham potential by comparison of the two integrands in the second
equation of (3.11),
~2 ³ mω ´h³ mω ´2 i
r
v K S (x) + ²11 = x −1
2m ~ ~
1 1
= mω2 x 2 − ~ω
2 2
= v ext − E 0 . (3.12)
Note that, as it is expected, the Kohn-Sham potential in absence of an inter-

particle interaction is indeed equal to the external field. Moreover, the Kohn-Sham
orbital is exactly the eigenfunction of the Hamiltonian. The Lagrange multiplier ²11
corresponds to minus the energy of the first orbital. Nevertheless, this is trifling, since
the potential is defined up to a constant.
In the one-orbital case the only unitary transformation we can apply to the KS
orbital is a change of sign. This is again unimportant, since this information is lost
in the density computation. The explicit formula for the solution of the inversion
problem in the one-orbital case [22] reads
p
∇2 ϕ1 (~
r ) ∇2 n(~
e r)
v K S (~
r)= = p (3.13)
2ϕ1 (~r) 2 n(~
e r)
The first test of the routine takes such density of a one-orbital unidimensional
quantum harmonic oscillator.
Based on the extent of the input density, we choose the compact space for the
minimization as [−8, 8], with a mesh h = 0.1. For the sake of simplicity we make use
of natural units, m = 1, ω = 1, and ~ = 1. We will restore those quantities to their
physical values while studying three-dimensional systems.
If we do not take any precaution, the test immediately exhibits the numerical
issues we have already discussed in the second chapter. In fact, while the minimizing
orbital is properly constructed in the entire space (see figure 3.1) almost indepen-
dently from the choice of the starting guess, figure 3.2 shows that the numerical
errors in the Kohn-Sham potential become important already for |x| > 1, where the
potential starts not to follow the expected parabolic behaviour. Instead, the shape of
the potential in the central region is reasonable.
If we make use of the one-orbital formula (3.13), instead of applying the CV
method, the potential is better reproduced, but still it stubles into numerical errors
in the outer regions, for |x| > 3.5; this is shown in figure 3.3.
It is interesting to point out that indeterminately refining the spatial grid does
not entail any sensitive improvement of the results; on the contrary, it does increase
the number of variables to be treated, and weights on the computational cost of the
routine.
The unsatisfactory results call for a new numerical approach; the scaling (2.29) of
the wave functions comes at hand. In fact, the results of the inversion are definitely
improved once we rescale all the terms in the cost functional. Such a remarkable
improvement is due to the fact that the rescaling process renders the quantity to be
minimized of the same order of magnitude in the entire space. This feature solves the
problem of roundings in the summation of wave functions characterized by a rapid,
exponential, decay to zero.
The program finds the right solution even for starting points that can be pretty
far from the expected solution, e.g. for strongly non-localized guesses such as a
0.8
KS Orbital
Harmonic Oscillator Eigenfunction
0.7
0.6
0.5
phi0(x)
0.4
0.3
0.2
0.1
0
-8 -6 -4 -2 0 2 4 6 8
x
Figure 3.1: The Kohn-Sham orbital for a single particle, subject to a harmonic oscillator trap,
is Gaussian and concide with the first eigenfunction of the harmonic oscillator.
2
KSP
Harmomonic Potential
1.5
1
vKS(x)
0.5
-0.5
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x
Figure 3.2: The Kohn-Sham potential as it results from a non-scaled one-orbital harmonic
oscillator inversion.
12
One-Orbital Formula
Harmonic Potential
10
6
vKS(x)
-2
-4 -2 0 2 4
x
Figure 3.3: The Kohn-Sham potential, computed through the one-orbital anlaytic for-
mula (3.13), compared with the expected potential.
sinusoidal wave function guess. The result is illustrated in figure 3.4, for a starting
guess of a constant scaled wave function.
In the present test of the routine we provided the analytical structure of the
first and of the second derivative of the density function, which are required by the
program in the scaled scheme (compare equation (2.17) and (2.31); derivatives of
the density appear in the latter), but the generalizing perspective brought us to a
numerical automation of those calculations in the following tests.
Many-Orbital Harmonic Oscillator
The major addition coming along with the increasing number of filled orbitals is
certainly the necessity of imposing orthonormality constraints.
p This is necessary also
to avoid trivial solutions of the inversion, e.g. φi = n(x)
e and φi 6= j = 0.
The insertion of a regularization scheme becomes more and more important, in
order to prevent oscillating solutions. In fact, the two different types of constraints
tend to compete in the search of the minimum, and may produce non-physical
results. The reason of this behaviour is that when more than one degree of freedom
is present, the inversion problem really becomes ill-posed. As we already discussed,
the insertion of a non-vanishing regularization term put then an artificial upper
bound to the norm of the gradient of the wave functions, it reduces the number of
possible solutions and renders them smoother. The effect of the presence of the
penalty functional, with a penalty parameter δ = 0.12 , is shown in figures 3.5 and 3.6
for the case of a two-orbital system. The effect is more marked if the number of filled
2 Equal to the mesh, that is our estimate of the numerical noise of the input density.
35
KSP
Harmonic Potential
30
25
20
vKS(x)
15
10
-5
-8 -6 -4 -2 0 2 4 6 8
x
Figure 3.4: The Kohn-Sham potential as obtained from a scaled one-orbital harmonic oscilla-
tor inversion.
orbitals in the system increases.

The increased number of degrees of freedom of the system results in a higher
number of possible unitary transformations of the wave functions to appear as
solutions of the minimization. In particular, the wave functions tend to mingle and
exchange one with each other, especially in the central region. Figure 3.7) shows
the Kohn-Sham wave functions obtained in a run of the software with three filled
orbitals, versus the eigenfunctions of a harmonic oscillator eigenvalue problem. It
is interesting to notice that the Kohn-Sham wave functions follow the parity of the
correspondent eigenfunctions. In other words, even (odd) Kohn-Sham orbitals are
unitary transformation of even (odd) eigenfunctions. No mixing is present. However,
those unitary transformations do not affect at all the Kohn-Sham potential structure;
figure 3.8 shows this aspect.
It also becomes important not to pick up a very poor starting guess; still, the
choice is quite free, once it respects some basic symmetries and properties of the
solution. In these tests we have used monomials (x/8) j 3 of rising degree as starting
point of the minimization; they approximately follow the parity of the solution, but
are satisfactorily far from it. It is necessary to be pretty specific in fixing the limits
of the compact space. The space must obviously be big enough to host orbitals that
respect the constraints. However, if such compact space is too big, the many trial
options to be explored may render the convergence unreachable.
The routine is in principle able to solve the inversion problem for an arbitrary
number of degrees of freedom, but with time and computational limits. In particular,
3 The guess is normalized in order not to go out from the compact space in which the software works.
0.8
Non-regularized KS Orbital
Regularized KS Orbital
0.7
0.6
0.5
phi0(x)
0.4
0.3
0.2
0.1
0
-6 -4 -2 0 2 4 6
x
Figure 3.5: The first Kohn-Sham orbitals resulting from a two-orbital inversion in absence
of a regularization scheme, compared with the same orbital calculated using the Tikhonov
regularization. The number of inflexion points diminishes.
0.8
Non-regularized KS Orbital
Regularized KS Orbital
0.6
0.4
0.2
phi1(x)
-0.2
-0.4
-0.6
-0.8
-6 -4 -2 0 2 4 6
x
Figure 3.6: The second Kohn-Sham orbitals resulting from a two-orbital inversion in absence
of a regularization scheme, compared with the same orbital calculated using the Tikhonov
regularization. The asymmetries disappear.
1
First KS Orbital
Second KS Orbital
Third KS Orbital
0.8 First HO Eigenfunction
Second HO Eigenfunction
Third HO Eigenfunction
0.6
0.4
0.2
phi0(x)
-0.2
-0.4
-0.6
-0.8
-1
-6 -4 -2 0 2 4 6
x
Figure 3.7: The Kohn-Sham orbitals resulting from a three-orbital inversion (solid lines)
compared with the first three eigenfunction of the harmonic oscillator (dashed lines).
14
KS Potential
Harmonic Oscillator Potential
12
10
8
vKS(x)
-2
-6 -4 -2 0 2 4 6
x
Figure 3.8: The Harmonic Kohn-Sham potential in a three-orbital system, as recovered from
the Kohn-Sham orbitals shown in figure 3.7.
450
Computational Time
2
en /60
400
350
300
250
t [s]
200
150
100
50
0
0 5 10 15 20
n [#Particles]
Figure 3.9: The computational time is a steep function of the number of filled orbital (degrees
of freedom of the problem) in the harmonic oscillator inversion. A qualitative comparison of
2
such function with e n /60 is also illustrated in the figure.
we point out that the computational time increases exponentially with the number
of degrees of freedom, and this is a great limitation especially when a poor starting
guess is chosen. Figure 3.9 exhibits this feature. The number of variables and the
amount of calculations done to evaluate the function in the routine are the main
reasons explaining this type of growth. Still, this is a harsh limit only if we would want
to use the software as a brute force tool, without taking into account the symmetries
and the properties of the problem under investigation. Instead, the perspective for
the usage of the program is good enough if we plan to face up problems in which
we are able to provide good starting points, and in which we have a knowledge of
features characterising the solution.
The spatial mesh is usually taken of the order of 10−1 fm, and the tolerance
for the acceptance of the solution is set to 10−4 fm, roughly what we expect to be
the numerical error correspondent to the chosen implementation of the numerical
methods. The above tolerance is set both as the relative error in order to ensure that
the iterations have effectively reached an optimal value and as the absolute value of
the constraints violation.
3.2.2 Morse Oscillator Density

Another group of tests deals with the density typical of a Morse Oscillator [47]
v ext (x) = D(1 − e −αx )2 , (3.14)
whose structure is depicted in figure 3.10. Variables x actually stands for r − r e , r e

being some equilibrium value. This potential is often used to describe the energy
30
10(1 - e- x/2)2
25
20
vext(x)
15
10
0
-2 0 2 4 6 8 10
x
Figure 3.10: The Morse Oscillator for D = 10 and α = 1/2.
spectrum of vibrating, non-rotating diatomic molecules. Also, it has been widely used
in the study of collision theory and intra- or inter-molecular bindings. For detailed
application of the Morse potential we address the reader to [48].
Our main interest in such potential, in comparison with the harmonic one, comes
from the fact that it saturates at infinity. The same happens for the nuclear potential,
that goes to zero at infinity. The system presents then a finite number of bound states
and its density gets certainly closer to those we will be meant to treat at the end of
the tests.
Before listing the eigenfunctions of a system subject to a Morse
p potential, we need
a brief series of definitions. Let us define ξ = e −αx and λ = 2D/α; let Γ(x) be the
Euler’s Gamma function, Z ∞
Γ(x) = d t t x−1 e −t , (3.15)
0
and L (α)
n (x) the so called associated Laguerre polynomials,
x −α e x d n −x n+α
L (α)
n (x) = (e x ), (3.16)
n! d x n
that satisfy the recurrence rule
nL α α α
n (x) = (−x + 2(n − 1) + α + 1)L n−1 (x) − (n − 1 + α)L n−2 (x) (3.17)
for n ≥ 2. Note that for α = 0, we recover the usual definition of the Laguerre polyno-
mials. Then, the system presents [λ + 1/2] normalized bound states, those that solve
the Schrödinger equation
~2 d 2 ψ
− + D(1 − e −αx )2 ψ = E ψ. (3.18)
2m d x 2
The explicit form of the eigenfunctions is found to be [49]
1 ξ
ψλ,n (ξ) = N (λ, n)ξλ−n− 2 e − 2 L (2λ−2n−1)
n (ξ), (3.19)
s
(2λ − 2n − 1)Γ(n + 1)
N (λ, n) = , (3.20)
Γ(2λ − n)
for n = 0, . . . [λ − 1/2].
We have chosen the potential parameters as α = 1/2, D = 10, so that λ ≈ 8.9 and
the potential well offers room for nine bound states. The inversion process behaves
for all of the possible degrees of freedom in the system. Here the interval of definition
of the orbitals is chosen as [−3, 6], with a mesh equal to h = 0.1, and penalty parameter
δ = 0.1. The starting guess for the minimization is a set of rising monomials, as
well as for the harmonic oscillator tests; this decision stems out from the fact that
the Morse potential in the neighbours of its minimum can be approximated to a
harmonic oscillator. Figure 3.11 exhibits the potential we have extracted for a seven-
orbital system. Some numerical noise is present in the intermediate part of the
potential. Those small oscillations, that can also be noticed in figure 3.8, have been
present in many results of the tests until, while analysing the three-dimensional
harmonic oscillator problem, we inserted a correction into the code, relatively to the
implementation of the Simpson’s integration method.
70
KS Potential
10(1 - e- x/2)2-4.5
60
50
40
vKS(x)
30
20
10
-10
-3 -2 -1 0 1 2 3 4 5 6
x
Figure 3.11: The Kohn-Sham potential from the density of seven-orbital in a Morse oscillator
trap.
3.3 Three-dimensional Test

Harmonic Oscillator Inversion
The first application of the three-dimensional spherical version of Constrained

Variational method has been done on a system characterized by the density typical
of an isotropic harmonic three-dimentional oscillator,
1
v ext (r ) = mω2 r 2 , (3.21)
2
1
where m = 939 MeV is the approximate mass of a nucleon and ω = 41/A 3 [24]. By
chosing those values for the mass and the frequency of the oscillator, we mean to
provide results that will be comparable, at least to some extent, to those we will obtain
with nuclear densities. In fact, as we have discussed while presenting the Nuclear
Shell Model, the harmonic oscillator is a first approximation of the nuclear average
potential that acts on the individuals nucleons, especially adequate in the inner part
of nuclei.
The solution of the Schödinger eigenvalue problem with an harmonic potential is
the set of eigenfunctions
2 (l + 1 )
ψnl m (r, θ, φ) = R nl (r )Yl m (θ, φ) = Nnl r l e −νr L n−12 (2νr 2 )Yl m (θ, φ), (3.22)
vs
u
t 2ν3 2k+2l +2 k!νl
u
Nnl = , (3.23)
π (2n + 2l − 1)!!
mω
ν= , (3.24)
2~
(l + 1 )
where L n−12 (2νr 2 ) are the generalized Laguerre polynomials already defined in equa-
tion (3.16). The corresponding eigenvalues,
³ 3´
E nl = ~ω 2(n − 1) + l + , (3.25)
2
set a clear energy order and define the degeneracy d = l (2l + 1) = (n+1)(n+2)
P
2 .
From the orbitals written in the decoupled basis, the density is obtained as
X 2l + 1
n(r ) = |R nl (r )|2 . (3.26)
(nl ) 4π
We can always decide how many particles to put in the system, up to some
numerical limit. The test was mainly developed to benchmark the application of
the inversion schemes to nuclei. Results are mainly positive and do not present
any new critical feature with respect to those we have already encountered in the
one-dimensional tests. We choose the interval of definition as [0, 12] fm, with mesh
h = 0.1 fm and penalty parameter δ = 0.1. The starting guess is taken as functions that
are proportional to the eigenfunctions of the harmonic oscillator. We report the result
of a succesful inversion, with the input of a four orbitals’ harmonic oscillator density
(see figure figure 3.12). In particular, the Kohn-Sham orbitals we obtain, shown in
figure 3.13, respect the number of nodes of the eigenfunction of the direct problem.
The Kohn-Sham potential relative to such density is illustrated in figure 3.14.
3.3. THREE-DIMENSIONAL TEST 57
n [fm-3]
0.06
0.05
0.04
0.03
0.02
0.01
0
0 2 4 6 8 10 12
r [fm]
Figure 3.12: The density of a system composed by twenty particles, distributed in the four
lowest-energy orbitals, in an isotropic harmonic oscillator trap.
]
]
-3/2
-3/2
0.2
phi [fm
phi [fm
0.4
0.18
0.35
0.16
0.3
0.14
0.25 0.12
0.2 0.1
0.08
0.15
0.06
0.1
0.04
0.05
0.02
0 0
0 2 4 6 8 10 12 0 2 4 6 8 10 12
r [fm] r [fm]
]
0.16
-3/2
-3/2
0.5
phi [fm
phi [fm
0.14
0.4
0.12
0.3
0.1
0.2
0.08
0.06 0.1
0.04 0
0.02
− 0.1
0
0 2 4 6 8 10 12 0 2 4 6 8 10 12
r [fm] r [fm]
Figure 3.13: The Kohn-Sham orbitals that are obtained by minimizing the objective function
highly resemble the 1s, 1p, 1d , and 2s eigenfunctions (respectively from left to right, top to
bottom) of the harmonic oscillator.
60
KS Potential
Harmonic Oscillator Potential
40
20
vKS(r) [MeV]
-20
-40
-60
0 1 2 3 4 5 6 7 8
r [fm]
Figure 3.14: The Kohn-Sham potential of a system composed by twenty particles, distributed
in the four lowest-energy orbitals, versus the expected result. The potentials are shifted to
provide a better qualitative comparison.
3.4. KS POTENTIAL FROM NUCLEAR DENSITIES 59
3.4 Kohn-Sham Potential Deduced from Nuclear Densi-

ties
With the aid of the Constrained Variational method we proceed to deduce the
Kohn-Sham potential from the experimental densities of magic nuclei. We have
already discussed, in the first section of the manuscript, how those nuclei present a
higher average binding energy than one would expect by analysing the semi-empirical
mass formula (1.3) and how they are more stable against decay with respect to other
configurations. Closed-shell systems are characterized by spherical symmetry, that
leads to a complete cancellation of the angular dependence of the quantities of
interest, the density (2.53) above all.
In the following, we will investigate four nuclei: 4 He, 16 O, 40 Ca, and 208 Pb. The
choice of studying each nucleus is guided by particular reasons;
4
• He is a one-orbital nucleus. We can then benchmark the results of our mini-
mization method against the one-orbital analytic and exact formula (3.13).
16
• O and 40 Ca are interesting cases of light and medium-mass nuclei. For each
of them, the literature makes available both the nSoG and the 3pF parametriza-
tions. We will especially focus on the investigation of model-dependence
features. This will hopefully help us to understand what are the limitations
of our procedure that do not depend on the physics of the system, but on the
implementation of our inversion routine.
208
• Pb is a heavy nucleus. We will address the study of this nucleus to propose
possible solutions to the model-dependence features that characterize exper-
imental densities. Moreover, the investigation of a theoretical Lead density,
together with other techniques, will be exploited to qualitatively understand
how errors propagate from the nuclear density to the Kohn-Sham potential.
We will organize the analysis of results concerning the study of each nuclear
potential from the lighter to the heavier nucleus. Our objective is to let the reader
understand which issues stem out when one deals with densities that are obtained
through a model-dependent analysis, other than being characterized by numerical
uncertainties. Also, the following exposition is thought in order to highlight the
differences between model-dependent and numerical issues, to clarify their causes
and to try proposing a solution to them.
3.4.1 The Helium Nucleus

The 4 He proton charge density is made available by [1] in the 10SoG parametriza-
tion. Figure 3.15 illustrates the proton density that has been used as the input of the
inversion scheme. The results of this inversion are an application of the numerical
method to a system that is characterized by one filled 1s 1/2 orbital, only. A compari-
son of the Kohn-Sham potential can be therefore done with the analytic, exact result
given by equation (3.13). The Helium nucleus represents a good laboratory in order to
begin comprehending the limitations that the choice of the density parametrization
has on the Kohn-Sham potential.
The Kohn-Sham potential of the numerical minimization coincides to that ob-
tained analytically. We can assert that the inversion procedure works fine in the
one-orbital case. In figure 3.16 we draw the result of the numerical inversion, only,
but the analytic potential would exactly coincide. Here and in the following in-
versions, the starting guess is the set of the eigenfunctions (3.22) of the harmonic
oscillator. The domain of the orbital is chosen as [0.1, 7] fm, with a mesh h = 0.1 fm.
Such values are chosen by looking at the input density, that can be considered to be
decayed to zero at the position of r ≈ 6 − 7 fm. We cannot explore the potential at
r = 0 fm, since the de-convoluted proton density (2.83) is not defined for such value.
The knowledge of the density does not carry any piece of information about the
value of the arbitrary constant of the Kohn-Sham potential, but since in the Helium
case an asymptotic behaviour appears evident, we can shift the result to let it go to
zero at large distance, as it should.
The potential has a reasonable depth of Vp ≈ −42 MeV, that is 5 MeV away
from the expected phenomenological value, according to the proton average po-
tential (1.28). The difference is about 10 %. The radial extent of the potential is
too long if compared to that expected, that is R(4 He) = 1.6755(28) fm 4 . If we look
at the position where the potential has reached half of its depth, this correspond
approximately to r = 3.5 fm, that is almost twice the expectation. Such spread may be
caused by the fact that we are considering a small finite nucleus within local density
approximation; contributions due to the gradient of the density function may be
decisive in determining the form of the Kohn-Sham potential.
If we extend the radial domain of the orbital to [0.1, 8], a non-physical behaviour
appears (see figure 3.17): the tail of the potential is wrong and resembles a parabolic
growth. This feature can be explained by the fact that the Gaussians used in the
nSoG parametrization of the density extend up to 4.9 fm. The tail of the density is
extrapolated as the tail of the last of the Gaussian of the parametrization. Indeed,
the potential that produces a Gaussian-modulated density is an harmonic oscillator
potential. We encounter here for the first time the strong limitation that characterizes
the nSoG parametrization. The functional assumption for the extrapolation of the
densities at large r in the experimental literature is found to be wrong.
This statement seems to collide with the results provided in [13], where apparently
positive and physically meaningful results have been deduced from a 12SoG Lead
neutron density of the Lead for the tail of the Kohn-Sham potential. We assert that
such result must be labelled as an inverse crime. In fact, the inversion method (vLB)
that has been used there is characterized by a great bias due to the fact that its starting
guess is given by a Woods-Saxon potential. Because of its particular convergence
criterion, the method does not modify the tail of the Woods-Saxon potential to let it
agree with the nSoG density. While discussing the case of Lead neutron density, we
will discuss a possible technique to proceed to a correct extrapolation of the tail.
4 https://www-nds.iaea.org/radii/
0.06
10SoG Density
0.05
0.04
n(r) [fm-3]
0.03
0.02
0.01
0
0 1 2 3 4 5 6 7
r [fm]
Figure 3.15: The 4 He proton density as obtained via a deconvolution from the proton charge
densy provided by De Vries [1].
5
KS Potential
-5
-10
-15
vKS(r) [MeV]
-20
-25
-30
-35
-40
-45
1 2 3 4 5 6 7
r [fm]
Figure 3.16: The Kohn-Sham potential deduced from the 4 He proton density.
50
KS Potential
40
30
20
10
vKS(r) [MeV]
-10
-20
-30
-40
-50
1 2 3 4 5 6 7 8 9
r [fm]
Figure 3.17: The Kohn-Sham potential, deduced from the 4 He proton density, presents a
non-physical parabolic growth at large r .
3.4.2 The Oxygen Nucleus

Oxygen is a light, doubly magic, nucleus, with eight protons and eight neutrons,
each of them filling the 1s 1/2 , 1p 3/2 and 1p 1/2 orbitals.
Literature [1] provides the proton density of Oxygen in two different parametriza-
tions (3pF and 12SoG), shown in figure 3.18. Their form suggests to choose the
interval of domain of the orbitals as [0.1, 8] fm, with a mesh h = 0.1 fm and a penalty
parameter δ = 0.1. Different parametrizations of the density lead to surprisingly
different densities:
• The 12SoG parametrization produces a density that is much more compact

than the 3pF. The radial extent of the 12SoG density is acceptable (R = 2.7 fm)
if compared to the charge radius of the Oxygen nucleus, that is R(16 O) =
2.6991(52) fm5 ; in contrast, the 3pF parametrization provides a wrong radius,
that is about 3.6 fm. However, as it can be seen from figure 3.19, the tail of
the latter parametrization is qualitatively correct (compare with the functional
behaviour (3.19) of the Morse oscillator density, that saturates at large r ), while
the tail of the 12SoG decays to zero too fast, as discussed for the case of Helium.
• A marked dip characterizes the 12SoG proton density. The value of the den-
sity at the origin is due to the contribution of the orbital 1s 1/2 , only, since the
centrifugal potential prevents other orbitals to explore the position r = 0. The
dip may then be caused by an emptying of the orbital 1s 1/2 in the target nu-
cleus. On the other hand, the density parametrized by a Fermi function is, by
construction, flat nearby the origin.
Figure 3.20 depicts the Kohn-Sham potential we have deduced from the 12SoG
proton density of Oxygen.
Non-physical oscillations stem out in the intermediate radial interval. Let us
recall that in scattering measurements the inner part of the density and its tail are
extrapolated. It seems that, since the radial extent of the nucleus is not very large, the
two critical radial intervals in which we extract the potential from the extrapolated
density do overlap, thus tainting the intermediate part of the potential, too. Moreover,
the potential seems to follow, at each point, the behaviour induced by the closest
Gaussian, only, and is therefore characterized by a sequence of parabolic behaviours.
This interpretation is enforced by the fact that the minimum of each parabola roughly
corresponds to the centres R i of one of the Gaussians in the parametrization.
The marked oscillation of the potential makes it impossible to determine the
physical asymptotic behaviour of the potential. We assume that the last parabola
with minimum at r = 6.5 fm is due to the last Gaussian of the parametrization (R i =
6.4 fm). Beyond this value, we say the result to be fully model-dependent. A rough
resemblance of the Kohn-Sham potential to a Woods-Saxon potential can then be
noticed for smaller radii.
The inversion scheme for the 12SoG density provides meaningful results only
in an intermediate radial interval, relatively to the depth (Vp ≈ −50 Mev) and to the
radial extent of the potential (R ≈ 3 fm).
Whenever a dip of the target density at the origin is present, the inversion scheme
reads it as a smaller probability for nucleons to occupy such position. The potential
deduced from the 12SoG density presents therefore a raise at small r that reflects
such physical interpretation.
0.09
3pF
12SoG
0.08
0.07
0.06
n(r) [fm-3]
0.05
0.04
0.03
0.02
0.01
0
0 1 2 3 4 5 6 7 8
r [fm]
Figure 3.18: The 16 O proton density as obtained via a deconvolution from the proton charge
density provided by De Vries [1] with a 3pF and a 12SoG parametrizations.
Figure 3.21 illustrates the Kohn-Sham potential deduced from the 3pF proton
density of Oxygen. The depth of the potential well is acceptably predicted. With
respect to the previous case, it is much easier to estimate it from the potential:
Vp ≈ −50 Mev. As we already noticed, the radial extent of the 3pF is wider than
in the previous case. Consequently, the same happens for the potential.
The non-physical oscillations that characterized the inversion with the 12SoG
density disappear in the 3pF case. This confirms our intuition; oscillations are caused
by the individual Gaussians that are used in the Sum of Gaussian parametrizations
and this effect is magnified in the density-to-potential inversion.
Except for some oscillations of the order of about 5 MeV, the asymptotic behaviour
of the tail of the potential is here clear and correct. The Fermi parametrizations are
characterized by a correct functional assumption for the behaviour of the tail of the
nuclear densities. Also, such feature points to the fact that our inversion method
extracts only the piece information that is carried by the experimental densities:
neither it stumbles upon inverse crimes, nor it generates physically meaningful
results on its own.
The potential at small r does not behave correctly and shows the same trend
as in the previous case. This must be then labelled as a limitation of our code. In
comparison with the 12SoG, the 3pF density has a lower value at the origin, but it
is flat. The correspondent raise of the Kohn-Sham potential is more marked than
before, but has no clear physical meaning.
0.10000000
3pF
12SoG
0.01000000
0.00100000
0.00010000
n(r) [fm-3]
0.00001000
0.00000100
0.00000010
0.00000001
0 1 2 3 4 5 6 7 8
r [fm]
Figure 3.19: The 16 O proton density as obtained via a deconvolution from the proton charge
density provided by De Vries [1] with a 3pF and a 12SoG parametrizations. Logarithmic scale is
used to highlight the functional behaviour of the tails.
30
KS Potential
WS Potential
20
10
-10
vKS(r)
-20
-30
-40
-50
-60
1 2 3 4 5 6 7
r [fm]
Figure 3.20: A comparison between the Kohn-Sham potential obtained by 16 O proton density
(12SoG) through the constrained-variational method and a Woods-Saxon potential.
vKS [MeV]
Kohn-Sham Potential
10
Woods-Saxon Potential
0
−10
−20
−30
−40
−50
0 2 4 6 8 10
r [fm]
Figure 3.21: Comparison between the Kohn-Sham potential obtained by 16 O proton density
(3pF) through the constrained-variational method and a Woods-Saxon potential.
3.4.3 The Calcium Nucleus

40
Ca proton density (figure 3.22) is provided by [1] in the 12SoG and 3pF parametriza-
tions. Calcium is a medium-mass magic nucleus, with twenty protons and twenty
neutrons, disposed in the six orbitals 1s 1/2 , 1p 3/2 , 1p 1/2 , 1d 5/2 , 2s 1/2 and 1d 3/2 . The
density of twenty particles in a harmonic oscillator trap is also shown in figure 3.22.
The box in which orbitals are defined is here [0.1, 9] fm, with mesh h = 0.1 fm and
penalty parameter δ = 0.1.
In contrast with the case of Oxygen proton densities, the two different parametriza-
tions appear to be more similar one another. However, it can be noticed that, again,
the 12SoG choice predicts a more compact distribution of protons than 3pF. The
12SoG density presents a raise at the origin, while 3pF is flat by construction.
The radial extent of the densities is in both cases (R = 3.8 fm for 3pF and R = 3.5 fm
for 12SoG) close to the expected value R(40 Ca) = 3.4776(19) fm 6 . With the aid of
figure 3.23, we remark again that the Sum of Gaussians density is more similar to an
harmonic oscillator density than to a Fermi function.
Figure 3.24 shows the Kohn-Sham potential obtained from the 12SoG parametriza-
tion. An equivalent harmonic oscillator potential is depicted, too.
By comparison with a Woods-Saxon potential (see figure 3.25), we understand
the the radial extent and the depth of the well are fairly well reproduced.
The non-physical oscillations in the intermediate radial interval of the potential,
that characterized the Oxygen case, have here disappeared. This confirms the fact
that oscillations were caused by the overlap of the two previously mentioned critical
radial regions.
The potential at low r does not behave properly, since the dip in the density
that was characterizing the Oxygen 12SoG density is here absent. The raise is not
physical as that of the Oxygen potential, but, again, it has no physical explanation.
The difference between the densities at the origin is remarkable. According to the
discussion we have done in the first chapter, we know that the interior of the nuclear
average potential can be well approximated by an harmonic oscillator potential. Also,
we know from the three-dimensional test we have done in the previous section that
0.14
3pF Density
12SoG Density
HO density
0.12
0.1
0.08
n(r) [fm-3]
0.06
0.04
0.02
0
0 1 2 3 4 5 6 7 8 9
r [fm]
Figure 3.22: The 40 Ca proton density as obtained via a deconvolution from the proton charge
density provided by De Vries [1] with a 3pF and a 12SoG parametrizations. The density of
twenty particles in an isotropic harmonic oscillator trap is also depicted.
our inversion method works fine in the harmonic oscillator case. It seems that our
inversion method misinterprets the that are values assumed by the experimental
densities at the origin: when they are lower than the harmonic oscillator density, the
inversion produces a non-physical raise that appears in many of our Kohn-Sham
potentials.
An equivalent (in the sense specified above) harmonic potential is shown in
figure 3.24; since such potential is an approximation of the nuclear average potential
up to the radial extent of the potential well, we can read out from the figure what
should be the expected form of the Calcium Kohn-Sham potential at the origin.
Figure 3.25 reports the Kohn-Sham potential deduced by the 3pF-parametrized
density. The form of the effective potential in the intermediate region [1.8, 5] fm is
similar to that of the potential deduced by the 12SoG density.
The raise of the potential here is much more pronounced.
An asymptotic behaviour at large r is clear. This confirm the adequate functional
assumption for the tail in 3pF densities.
1.000000
3pF Density
12SoG Density
HO density
0.100000
0.010000
n(r) [fm-3]
0.001000
0.000100
0.000010
1 2 3 4 5 6 7
r [fm]
Figure 3.23: The 40 Ca proton density as obtained via a deconvolution from the proton charge
density provided by De Vries [1] with a 3pF and a 12SoG parametrizations. The density of
twenty particles in an isotropic harmonic oscillator trap is also depicted. Logarithmic scale is
used to highlight the functional behaviour of the tails.
20
KS Potential
HO Potential
10
-10
vKS [MeV]
-20
-30
-40
-50
-60
1 2 3 4 5 6 7 8
r [fm]
Figure 3.24: Comparison between Kohn-Sham potential deduced by the experimental 12SoG
proton density of 40 Ca and that obtained from the density of twenty particles in an isotropic
harmonic oscillator trap.
vKS [MeV]
Kohn-Sham Potential
0
Woods-Saxon Potential
−10
−20
−30
−40
−50
0 1 2 3 4 5 6 7 8
r [fm]
Figure 3.25: Comparison between the Kohn-Sham potential obtained by 40 Ca proton density
(3pF) through the constrained-variational method and a Woods-Saxon potential.
3.4.4 Lead Nucleus

Lead is a heavy nucleus, with 82 protons and 126 neutrons. For the 22 orbital
occupancy, we refer to figure 1.3.
Based on what we have discussed up to now, we expect the Kohn-Sham potential
of 208 Pb to be more well-behaved than those of lighter nuclei, although its computa-
tional costs are higher. In fact, the radial extent of such nucleus, that is R 208 Pb = 7.4 fm
according to formula (1.2), is considerably bigger than in the other cases and we
expect the two previously discussed critical radial intervals to stay well separated.
The literature makes available both the neutron [2] and the proton [1] density; we
will limit our discussion to the former, because the absence of Coulomb effects will
be likely to help us to better understand the inversion issues. Figure 3.26 shows the
neutron density of Lead, as obtained by experimental measurements in the 12SoG
parametrization. Again, the functional assumption for the tail of the density diverges
from the expectations (see figure 3.27). In this figure we also report a theoretical
density obtained via Hartree-Fock calculations; we will come back to this below. We
choose the radial domain as [0, 12] fm, with mesh h = 0.1 and δ = 0.1.
Figure 3.28 shows the Kohn-Sham potential as deduced from the 12SoG neutron
density, compared with a Woods-Saxon potential.
The interior part of the potential shows oscillations of amplitude of the order of
5 MeV; we ascribe those to similar reasons to those we have discussed while analysing
the case of Helium.
The depth and the radial extent of the potential are slightly smaller than expected
(Vn ≈ −45 MeV and R ≈ 7 fm), while the behaviour of the tail is, as expected, wrong.
Since we have clearly established that the tail divergence is caused by the wrong
asymptotic of the experimental density parametrization, we are justified to extrapo-
late, through a fit, a more reasonable form of the tail. The idea is to search for the first
inflexion point of the potential after the centre of the last Gaussian of the parametriza-
tion; in the case of Lead, this is at R i = 8.7 fm. We assume that such inflexion point
is the indication of where the tail of the density starts being extrapolated in a non-
physical way. We substitute the tail of a Woods-Saxon potential (see equation (1.28))
to the divergent tail we have found. The depth of the Woods-Saxon potential is set to
V0 ≈ 50 MeV, in agreement with the phenomenological expectation; we impose the
two potentials to have the same value at the inflexion point r ∗ (continuity),
V0
v K S (r ∗ ) = − r ∗ −R
+ ², (3.27)
1+e a
and their first derivative to be equal there, too (differentiability),

r ∗ −R
V0 e a
v K0 S (r ∗ ) = ∗ . (3.28)
a (1 + e r a−R )2
The two equations above determine R[V0 , ²] and a[V0 , ²]. Finally, we fit the value of
the arbitrary constant by assigning to it a value that makes the match of the potentials
qualitatively as smooth as possible. A lower bound for ² is given by the value of the
highest energy ² j of the Kohn-Sham orbitals, that can be obtained by diagonalizing
the matrix ² j k of the orthonormality Lagrange multipliers. The potential with the
corrected tail is shown in figure 3.29.
0.12
HF Density
Experimental Density
0.1
0.08
n(r) [fm-3]
0.06
0.04
0.02
0
0 2 4 6 8 10 12 14 16 18 20
r [fm]
Figure 3.26: The neutron Lead density provided by Zenihiro [2] compared with the density
obtained by solving a direct problem within Hartree-Fock approximation, in which a Skyrme
force, SkP [3], is assumed.
1.0000000
HF Density
0.1000000
0.0100000
0.0010000
0.0001000
0.0000100
0.0000010
0.0000001
2 4 6 8 10 12
Figure 3.27: The neutron Lead density provided by Zenihiro [2] compared with the density
obtained by solving a direct problem within Hartree-Fock approximation, in which a Skyrme
force, SkP [3], is assumed. Logarithmic scale is used to better analyse the tails of the densities.
80
KS Potential
WS Potential
60
40
20
vKS [MeV]
-20
-40
-60
0 2 4 6 8 10 12
r [fm]
Figure 3.28: The Kohn-Sham potential as deduced from the experimental neutron density,
compared with a Woods-Saxon potential.
0
KS Potential
WS Potential
-5
-10
-15
-20
vKS [MeV]
-25
-30
-35
-40
-45
0 2 4 6 8 10 12
r [fm]
Figure 3.29: The Kohn-Sham potential as deduced from the experimental neutron density,
with a corrected Woods-Saxon tail.
Qualitative Analysis of the Propagation of the Errors
We have defined a method for the correction of the model-dependent behaviour

of the tail of the Kohn-Sham potential derived from a nSoG-parametrized density.
Thus, we proceed to a qualitative analysis of the errors that characterized the inver-
sion method itself.
In addition to what we have already shown, we studied the Kohn-Sham potential
obtained from a theoretical density provided by Hartree-Fock calculations7
The theoretical density is depicted in figure 3.26 and 3.27. There is indeed a
qualitative spread between the experimental and the theoretical density, that is
especially visible (≈ 30%) at the origin.
In order to understand if our correction to the tail of the Kohn-Sham potential is
satisfactory, we study the inversion problem relatively to a density that is theoretically
calculated. The tail of the latter density is indeed correct (see figure 3.27), since it
is produced by a potential that go to zero correctly. A benchmark against which
our corrected Kohn-Sham potential 3.29 can be compared is therefore available.
Figure 3.30 illustrates the comparison between the potential obtained by the Hartree-
Fock theoretical density, the Kohn-Sham potential obtained by the experimental
density, and the original potential that has been used in HF calculations of the
density. Results are in agreement in the central radial interval, let us say from 2 fm to
8.5 fm; the tails are in acceptable agreement only if we average the oscillations with
amplitude of 5 MeV that characterize the Kohn-Sham potential obtained from the
HF density at its large r interval. Although we have seen this is the typical magnitude
of numerical noise, such feature does not let us understand the correct asympotic
behaviour of the potential. In general, the absence of any clear asymptotic trend,
combined with the fact that potentials are defined up to an arbitrary constant, renders
the comparisons possible only at a qualitative level. The behaviour at the origin is
clearly wrong, since it shows a non-physical growth. Again, the depth and the radial
extent are compatible.
Another way to understand the level of correctness of our inversion method is
given by a comparison with the results of another inversion scheme. Thus, we can
fully understand which are the limitations of the input data and which are instead
the limitations due to the implementation of a general algorithm to solve the inverse
Kohn-Sham problem. In figure 3.31 we compare the Kohn-Sham potentials resulting
from our Constrained Variational Method and the van Leeuween and Baerends
method. The figure reports the potential assumed in Hartree-Fock calculations, too.
It is remarkable to note that the two Kohn-Sham potentials envelop the theoretical
potential. Two observations are in order.
• The results of the vLB method are highly dependent from the initial guess for
the potential. Here, such starting guess is taken as a Woods-Saxon potential,
with an adequate choice of the parameters.
7 A very successful theoretical approach to nuclear structure is the self-consistent mean-field method.
The approach is based on an independent particle picture, where nucleons are considered to be self-
bound by the average of the two-body interactions over the states occupied by the other particles. The
resulting field is created in a self-consistent way. Such a field is considered to be static, so that dynamical
corrections are neglected. Specifically, this approach can be realized by means of an adopted effective
interaction, solved at the level of the Hartree-Fock approximation. In the present case, the calculations
have been carried out within a Skyrme-type force, SkP [3], in absence of a Coulomb term. The potential
is used to solve the Schrödinger eigenvalue problem in Hartree-Fock approximation, in order to provide
eigenfunctions that are used to compute the density of the nucleus.
• In the inner region, the spread between the results obtained with the two differ-
ent methods is ≈ 25 MeV, that is about 50 % of the depth of the potential. Such
spread diminishes to the value of ≈ 5 MeV (10 % of the depth) for intermediate
radial coordinates, let us say 1.5 ≤ r ≤ 8.5 fm.
A final possibility for understanding the level of the uncertainties in the density-
to-potential inversion is to solve the direct problem with our Kohn-Sham potential.
We do this by using the potential we retain to be more similar to the true solution of
the corresponding inverse problem: the one deduced from the Lead experimental
neutron density (12SoG-parametrized), with the corrected tail. We compare the den-
sity thus obtained with the input density of the inverse problem. A naive expectation
would be to find exactly the density we have provided to the inversion method as an
input. Clearly, this is not what happens, since the entire inversion plus direct process
generates a certain amount of numerical noise in the quantities of interest. There is a
main advantage in comparing densities instead of potentials. In the latter case, we
must assume some asymptotic behaviour (the arbitrary constant of the potential) in
order to be able to superpose the potentials and provide a quantitative comparison.
Such problem is absent while contrasting densities. Figures 3.32 and 3.33 depict
the original experimental neutron density and the one that we have deduced from
the corresponding Kohn-Sham potential. Differences are evident, and 3.34 plots the
relative spread between the two of them; at the origin the difference is, as we have
already qualitatively estimated right above, around 40 %; in the intermediate region
the spread decreases down to 5 %. The spread in the tail cannot be compared, since
we have inserted an artificial correction to the tail of the potential.
It seems that, within the inversion process, the calculation of the Kohn-Sham
orbitals is non-problematic; on the other hand, when one tries to deduce from those
the form of the Kohn-Sham potential, the numerical uncertainties get incredibly
magnified.
10
KS Potential from HF Density
KS Potential from Exp Density
HF Potential
-10
-20
vKS(r) [MeV]
-30
-40
-50
-60
2 4 6 8 10 12
r [fm]
Figure 3.30: The Kohn-Sham potentials deduced from the Hartree-Fock theoretical and from
the experimental neutron densities, compared with the potential assumed in the Hartree-Fock
calculations of the density.
0
KS Potential with CV Method
KS Potential with vLB Method
HF Potential
-10
-20
vKS(r) [MeV]
-30
-40
-50
-60
0 2 4 6 8 10 12
r [fm]
Figure 3.31: The Kohn-Sham potential deduced from experimental neutron density through
two different inversion schemes, the CV method and the vLB method, compared with the
potential assumed in the Hartree-Fock calculations of the density.
0.12
Direct Problem Solution
0.1
0.08
n(r) [fm-3]
0.06
0.04
0.02
0
0 2 4 6 8 10 12
r [fm]
Figure 3.32: The density that is obtained by solving the Kohn-Sham direct problem using the
Kohn-Sham potential deduced by experimental neutron density, compared with the experi-
mental neutron density.
1
Direct Problem Solution
0.1
0.01
0.001
n(r) [fm-3]
0.0001
1e-05
1e-06
1e-07
0 2 4 6 8 10 12
r [fm]
Figure 3.33: The density obtained by solving the Kohn-Sham direct problem using the Kohn-
Sham potential deduced by experimental neutron density compare with the experimental
neutron density. Logarithmic scale is used to analyse the tails of the densities.
1
Relative Spread
0.9
0.8
0.7
0.6
|nKS-nexp|/nexp
0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12
r [fm]
Figure 3.34: The relative spread between the density obtained by solving the Kohn-Sham
direct problem using the Kohn-Sham potential deduced by experimental neutron density and
the experimental neutron density.
Conclusions
The possibility of testing an exact functional through the knowledge of the exact
effective potential that it should generate, is worth being investigated. This is espe-
cially true in the field of Nuclear Physics, where theoretical models are mostly based
on phenomenology. The Hohenberg and Kohn DFT is indeed exact, but the ingre-
dients of which the energy density functional consists cannot be obtained without
relying on drastic approximations. In contrast, the construction of the exact Kohn-
Sham potential from the nuclear density is in this sense promising and pioneering
in Nuclear Physics, since it would provide a reliable benchmark against which the
effective potentials assumed in energy density functional available on the market can
be tested.
The original project of this thesis was to develop a numerical method to address
the solution of the inverse Kohn-Sham problem in Nuclear Physics in the most general
case. However, while encoding the first unidimensional tests, we have realized that
the computational time necessary for calculations was a steep function (see figure 3.9)
of the number of degrees of freedom in the problem. This feature suggested us to first
proceed to the the application of our method on systems characterized by spherical
symmetry. To this purpose, we have defined a new formalism for the Constrained
Variational method, that has been presented in section 2.5.2. This formalism clearly
sets a limitation to the range of the physical systems to be subject of investigation,
but it follows the general idea to proceed step-by-step in the implementation of the
inversion method.
The study of the Kohn-Sham potentials, as deduced by the experimental densities
of a certain number of magic nuclei, has been treated in the bachelor’s theses [12], [13]
and [14], with a various level of detail and with different results. In all those works,
however, a simple inversion method (the van Leeuwen and Baerends’) has been used.
As we have pointed out in this manuscript (in sections 3.4.1 and 3.4.4), such iterative
method is highly biased by the initial guess on the potential, especially for which
regards the behaviour of the tail of the Kohn-Sham potential. Our perspective has
then been to build a method affected as least as possible by the initial assumptions
on the solution of the non-linear inverse problem. We believe that the choice of our
inversion method is particularly adequate to this aim.
The Constrained Variational method explores a much wider space in its search
for the solution than what happens for iterative approaches. However, this somehow
also represents a drawback of our implementation. In fact, whenever the input
densities present some type of error or model-dependence, our scheme is not able
to produce meaningful physical results. This is the case, for instance, for the tails of
the experimental densities parametrized by nSoG, that are characterized by a wrong
functional assumption. Moreover, our output is sensitive to small variations of the
input density. Densities that should, at least in principle, be related to the same
effective potential, wind up in different results. This is what happens to our nuclear
Kohn-Sham potential at small radii, where slight modifications of the occupancy of s
orbital entail the appearance of non-physical behaviours. Since our procedure does
not rely on a strong guidance towards the correct solution, such as a very valid initial
guess for the potential, it does not produce physical results in some critical radial
intervals, and there it is sensitive to numerical noise.
The application of the inversion Constrained Variational method on analytic
systems has given acceptable results in all the cases. Such a wide testing procedure
of the method represents itself a first objective of the thesis that we can consider fully
accomplished.
On the other hand, we encountered a surprising amount of troubles when we
considered as the input of the inversion process realistic densities, characterized by
systematic physical uncertainties. The results we have presented in section 3.4 are
not satisfactory. Although a partial justification is given by the high uncertainties
and model-dependence that characterize the inner and the outer regions of the
experimental nuclear densities, our method is not robust in those intervals, even
when nuclear densities obtained theoretically are used. However, we believe that the
sources of such non-physical behaviours of the nuclear effective potentials have been
clearly identified in most of cases. Unluckily, we could not always find a practical and
theoretically sound correction to those issues.
Further work must be done in this sense, and we have pointed to a clear possible
direction for improvements: the exploitation of the different methods to address
the solution of the inverse problem. In fact, none of the methods available on the
market seems to be free of drawbacks that jeopardize the possibility of providing solid
solutions to the Kohn-Sham inverse problem. For instance, our method could be
used to provide an unbiased solution to the inverse Kohn-Sham problem; the parallel
application of an iterative method could help instead to get physical corrections
in those intervals where our solution is tainted by the uncertainties and model-
dependence of the input density.
As a final observation, we would like to remark that our method is explicitly
prepared in order to be generalized to more complex systems. The computational
costs of each particular numerical method that is used in the inversion scheme
represents the main obstacle to such generalizations. One could then think to face
the inverse problem in a more generic way, without relying on assumptions about
the symmetries of the problem.
Appendix A
The Method of Lagrange

Multipliers
In general, some issues arise whenever one tries to perform a constrained multi-
dimensional optimization. Just to mention the most trivial of those complications,
derivatives cannot be defined properly on the borders of the regions in which the
problem is defined, if those regions are not open spaces.
Let us consider a functional F [ f ]. A necessary condition for some function to be
an extremum of such functional is that the variation of the functional vanishes for all
δ f (x),
δF
Z
δF = d x δ f (x) = 0 (A.1)
δ f (x)
It is then equivalent to require that the optimal solution satisfies
δF
= 0. (A.2)
δ f (x)
Also, if one keeps in mind the particular case in which the functional is an action
Z
F [ f ] = d x L ( f , f 0 ; x) = 0, (A.3)
then the condition (A.2) produces the well known Euler-Lagrange equations.
If some constraints are imposed to the optimization problem, such as
G i [ f ] = 0, (A.4)
it is often convenient to make use of the method of Lagrange multipliers. The method
consists in building an auxiliary quantity, the so-called Lagrangian, defined as
Λ[ f , λi ] = F [ f ] + λi G i [ f ] = 0.
X
(A.5)
i
It is possible to proof that this new problem, namely the free optimization of the
Lagrangian with respect to f and to the Lagrange multipliers λi is equivalent to
the original constrained problem. Thereby, one finds a candidate extremal of (A.5),
which will be parametrized by those λi , and afterwards imposes the constraints (A.4),
fixes the values of the multipliers. Note that the method provides only a necessary
81
82 APPENDIX A. THE METHOD OF LAGRANGE MULTIPLIERS
condition for functions to be an extremal of the functional investigated. Further

information is required in order to understand if the extrema one finds are indeed
minima, maxima or saddles. For more details on the method of Lagrange multipliers
we address the reader to [50, 51].
Appendix B
Density Operators and

Scalar/Vector Densities
B.0.1 Density Operators

One can consider a more fundamental definition of the density function, based
on density operators. The quantity
Ψ A (~ x A )Ψ∗A (~
x 1 , . . . ,~ x 1 , . . . ,~
xA) (B.1)
represents the probability distribution associated with a particular solution of the

Schrödinger equation. Consider now quanties such as
γ A = Ψ A (~
x01, . . . , ~
x 0 A )Ψ∗A (~
x 1 , . . . ,~
x A ). (B.2)
It is natural to construct a density matrix, whose elements are the γ A . The diagonal
entries of the matrix are indeed the original (B.1). (B.2) can be also thought as the
representation in the coordinates basis of the density operator
γ̂ A = |Ψ A 〉 〈Ψ A | . (B.3)
The density operator is a projection operator, and

Z
xdA ~
tr(γ̂ A ) = d A ~ x 0 Ψ A (~
x01, . . . , ~
x 0 A )Ψ∗A (~
x 1 , . . . ,~
xA) = 1 (B.4)
if Ψ A is normalized. Moreover, it is trivial to show that
〈 Â〉 = tr(γ̂ A Â). (B.5)
for each observable Â. The previous equation states a one-to-one mapping between
the density operator and the obsevables of the system under investigation. In fact,
by comparison with (1.7), one notices that the density operator does carry the same
piece of information as any A-nucleon wave function |Ψ A 〉. Furthermore, while |Ψ A 〉
is defined up to an arbitrary phase factor, γ̂ A is unique. As well as each operator
associated to some observable must be, γ̂ A is Hermitian, too.
The operator-like description becomes essential whenever the system under
investigation is part of some enviroment, and it is not isolated from it. Indeed, in such
a case the system does not have a complete Hamiltonian containing all its degrees
of freedom. Then, the system is open, and this feature precludes any possibility of a
wave function description [52].
83
84 APPENDIX B. DENSITY OPERATORS AND SCALAR/VECTOR DENSITIES
B.0.2 Reduced Density Matrices

We are mainly interested in one- or two-particle operators operators. Let us
recall that wave functions are antisymmetric. These two features allow to simplify
the density matrix, to produce a reduced density matrix or, for a spin-compensated
sysyem, a spinless density matrix. Let us call (B.1) A t h -order density matrix. One then
defines the reduced density matrix of order p through
γp (~
x01, ~
x02 . . . ~
x 0 p ;~
x 1 ,~
x 2 . . .~
xp )
Ã !Z
A
Z
= d~x p+1 · · · d~ x A γ A (~ x01, . . . ~
x 0 A ;~
x 1 , . . .~
x A ). (B.6)
p
In particular, second-order density matrix

A(A − 1)
Z Z
~ ~
γ2 (x 1 , x 2 ;~
0 0 x 1 ,~
x2 ) = d~ x A γ A (~
x 3 · · · d~ x01, . . . ~
x 0 A ;~
x 1 , . . .~
xA) (B.7)
2
integrates to the number of nucleon pairs A(A−1) 2 , while the first-order density matrix
Z Z
γ1 (~
x 0 1 ;~
x 1 ) = A d~ x A γ A (~
x 2 · · · d~ x01, . . . ~
x 0 A ;~
x 1 , . . .~
xA) (B.8)
normalizes to the number of nucleons A. These operators are positive semi-definite

and Hermitian, as well as γ̂ A . Also, antisimmetry requires that interchange of two
primed or two unprimed indices brings along a change of the sign of the density
operators.
B.0.3 Spinless density matrices

It is natural, whenever the quantities of interest do not involve the spin coordi-
nates, to further simplify the density matrices by a summation over σ1 (and σ2 ). The
first-order and the second-order spinless density matrices read
Z
n 1 (~ r 1 ) = d σ1 γ1 (~
r 0 1 ;~ x 0 1 ;~
x 1 ), (B.9)
Z Z
n 2 (~
r 01, ~r 0 2 ;~
r 1 ,~
r2) = d σ1 d σ2 γ2 (~
x01, ~
x 0 2 ;~
x 1 ,~
x 2 ), (B.10)
(B.11)
and the two of them are linked by the relation

2
Z
n 1 (~
r 0 1 ;~
r1) = r 2 n 2 (~
d~ r 0 1 ,~
r 2 ;~
r 1 ,~
r 2 ). (B.12)
A −1
Note that the diagonal entries of n 1 (~ r 0 1 ;~
r 1 ) provide just the nuclear density (1.29).
One can write down an energy formula
Z h ~2 Z
∇~2r n 1 (~
i
E = d~ r − r 0 ;~
r ) ~0 + d~ r v(~
r )n 1 (~
r)
2m r =~ r
Z Z
+ d~
r 1 d~
r 2 w(~ r 1 ,~
r 2 )n 2 (~r 1 ,~r 2 ). (B.13)
The three terms represent respectively the kinetic term, the external potential energy,
and the nucleon-nucleon interaction energy. The main advantage of such formula is
the fact that it involves only a function of three coordinates, n(r ), and two functions
of six coordinates, namely n 1 (~r 0 ;~
r ) and n 2 (~
r 1 ,~
r 2 ).
85
B.0.4 Scalar and Vector Densities

All along the thesis the term density means barionic density. In general, density
functional theory, and therefore the inverse Kohn-Sham problem, other than being
applicable to generic barionic system, can be considered in relation to other types of
densities. In [53] and [54], a systematic construction of the energy-density functional,
within local density approximation (LDA), is presented. If one considers the descrip-
tion of systems in which the time-reversal symmetry is broken, e. g. fast rotating
nuclei [55], the density matrix and the resulting mean-field will be characterized by
both time-even and time-odd components. While properties of time-even mean-
field are known rather well, also thanks to experiments, the knowledge of time-odd
mean-fields is fewer. The density of nuclear matter in the interior part of nuclei has
a well known definite value, namely the saturation density, which is to some extent
independent of the nuclear size. Because of that, a basic approximation is to consider
the state of matter at some point in the nucleus as it would be given in in infinite
matter at the corresponding density. In other words, the influence of different points
of nuclear matter on properties that are calculated at some given point is rather little.
It is such observation that justifies the adoption of local density approximation in
nuclear systems and in other fields of Physics.
The finite nuclear size implies that corrections must be taken into account for the
simplest version of LDA. Those corrections are implemented by local derivatives of
the density. The total energy of the nucleus, in LDA, is given by the integral of a local
energy density Z
E= d 3~
r H (~
r) (B.14)
and the energy density H depends on the one-body density matrix and on its deriva-
tives, usually up to the second order.
The type of nucleon considered is accounted via the isospin degree of freedom; if
we neglect it, the density matrix reads (compare the following with (B.8))
n(~ r 0 σ0 ) = 〈Ψ|a † (~
r σ,~ r 0 σ0 )a(~
r σ)|Ψ〉 . (B.15)
One can separate the spin degrees of freedom by defining a scalar and a vector part
of the density matrix, respectively
r 0 ) = n(~ r 0 σ0 )
r σ,~
X
n(~
r ,~ (B.16)
σ
~ r 0) = r 0 σ0 ) 〈σ0 |~
r σ,~ σ|σ〉 .
X
s(~
r ,~ n(~ (B.17)
σσ0
The density matrix is hermitian, and therefore its scalar and vector parts are sym-
metric under the exchange of the spatial arguments. Instead, the density matrix
corresponding to time-reversed states read
n T (~ r 0 ) = n † (~
r ,~ r 0 ) = n(~
r ,~ r 0 ,~
r) (B.18)
s T (~
~ r ,~ s † (~
r 0 ) = −~ r 0 ) = −~
r ,~ r 0 ,~
s(~ r) (B.19)
That means that the scalar density matrix is a real, symmetric, function of the spatial
arguments. On the other hand, the vector density matrix is imaginary and antisym-
metric. The local scalar and vector densities,
n(~
r ) = n(~
r ,~
r ), (B.20)
~
s(~
r ) =~
s(~
r ,~
r ), (B.21)
86 APPENDIX B. DENSITY OPERATORS AND SCALAR/VECTOR DENSITIES
respectively do not and do change sign under time reversal. In other words, the
barionic density is time-even, while the spin density is time-odd. In particular, one
acts on the scalar and vector matrices with the operators (∇ − ∇0 )/2i , and afterwards
sets ~
r =~r 0 . At the first order one obtains the current density ~
j and the spin current
density J µν (~r ). In the second order one applies the derivation above once again
and gets the kinetic density and the spin kinetic density. According to some well-
established rules, one constructs the local energy density as a sum of terms depending
on the different densities just defined. In particular, each term must be quadratic in
the local density; terms beyond the second order must be neglected; the energy must
be invariant with respect to parity, time reversal, and rotations. By analysing all of
such possible terms, it turns out that each density gives rise to one of the two main
terms of the interactive part of energy density:
~2
H (~
r)= τ0 + H even (~
r ) + H odd (~
r ). (B.22)
2m
The scalar class of densities gives rise to the first interaction term, while the vector
density and its local derivatives produce the second one.
To take into account the isospin quantum number, that is to consider differ-
ent types of nucleons, the expression above must be labelled as H t (~ r ), with t = 0
standing for the isoscalar energy density and t = 1 for isovector one. Each of those
quantities derives, respectively, by some n 0 or n 1 . Those are defined as the sum and
the difference of the proton and neutron contributions:
n0 = nn + n p (B.23)
n1 = nn − n p , (B.24)
and similar expression can be used to define other densities.

A variation of the energy density with respect to the possible six density functions
produces mean fields Γeven,odd
t , and the neutron or proton single-particle Hamiltoni-
ans can be obtained by combining the kinetic energy with the mean fields:
~2 ∇2
h n,p = − + Γeven
0 + Γodd even
0 ± (Γ1 + Γodd
1 ) (B.25)
2m n,p
Appendix C
Time-Independent Systems
As the time derivative in the Schrödinger equation appears linearly, one normally
must consider complex fields [56] (or wave functions in the first-quantized picture).
Those are obtained by the superposition of two independent real fields,
ψ1 + i ψ2
ψ= p , (C.1)
2
ψ1 − i ψ 2
ψ† = p (C.2)
2
The Lagrangian density of a Schrödinger field is
~2
L = i ψ† ψ̇ − ∇ψ† · ∇ψ − ψ† V (~
r , t )ψ, (C.3)
2m
whence Euler-Lagrange equations
δL δL ³ δL δL ´†
−∇· = −∇· =0 (C.4)
δψ δ∇ψ δψ δ∇ψ
return the Schrödinger equation and its complex conjugate
~2
− ∇2 ψ† + V ψ† = −i ψ̇† , (C.5)
2m
~2
− ∇2 ψ + V ψ = i ψ̇. (C.6)
2m
The conjugate variables are
δL
π= = i ψ† ,
δψ̇
δL
πT = = 0, (C.7)
δψ̇†
so that the Hamiltonian density reads
~2
H = πψ̇ + πT ψ̇† − L = −i ∇π · ∇ψ − i πV ψ. (C.8)
2m
87
88 APPENDIX C. TIME-INDEPENDENT SYSTEMS
This leads to the usual Hamiltonian of a Schrödinger field,

Z Z h ~2 i
H= rH =
d~ r ψ† −
d~ ∇2 + V ψ (C.9)
Ω Ω 2m
If the fields are time-independent, the right hand side of Schrödinger equations (C.5)
and (C.6) vanishes and the adjoint equation becomes trivial.
Appendix D
Deconvolution from Proton

Charge Density to Proton
Density
It is a matter of fact that experimental measurements are often limited to provide

proton charge density distributions in nuclei, as it happens in De Vries [1], instead of
nuclear densities. The reason is that in most of the experiments cross section data
are obtained by using charged probes, such as electron beams, which does interact
with protons, only.
Thus, the natural question that arises is how to move from some charge density
distribution to the respective particle density distribution. The answer is rather
simple, but it demands for some calculations.
Suppose one knows some radial proton charge density n ch (r ). Then, the proton
density n(r ) will be linked to the proton charge density by a convolution:
Z
0 0
n ch (r ) = [F proton (r − r ) ∗ n(r )](r ) = d r 0 F proton (r − r 0 )n(r 0 ), (D.1)
where F proton (r − r 0 ) is the form factor of a proton with respect to the electric force.
Such quantity represents the shape of the proton, which is not considered a point,
as it would be deduced by the cross section obtained using beams that interact with
it through the electric force. The form factor must normalize to the proton charge
q e . Therefore, it is necessary to apply a deconvolution on the above equation to
reverse it with respect to n(r ). The key idea is that convolution operations are much
more natural to be performed in the momentum space, where they correspond to
the product of the Fourier transforms of the two functions involved,
n̂ ch (q) = F̂ proton (q)n̂(q). (D.2)
Once n̂(q) has been calculated, it is straightforward to get n(r ) via a backward Fourier
transformation
n(r ) = (n̂(q))∨ (D.3)
As an example of such a deconvolution, consider a proton charge density given

by a sum of Gaussians (nSoG). Here the parametrization is used just as an example to
89
90APPENDIX D. DECONVOLUTION FROM PROTON CHARGE DENSITY TO PROTON DENSITY
show how to apply the deconvolution procedure.
(r −R i )2 (r +R i )2
X Ai h
γ2 γ2
i
n ch (r ) = 3
e +e , (D.4)
i 2π 2 γ3
where
q e ZQ i
Ai = . (D.5)
R2
1 + 2 γ2i
P
Q i and R i are parameters and i Q i = 1, Z is the atomic number of the nucleus, and
γ is a width parameter proportional to the rms radius of the nucleus. Let us recall
some useful formulas, fundamental in order to perform the calculations below. First,
in spherical coordinates the Fourier transformation and anti-transformation read
sin(qr )
Z ∞
fˆ(q) = 4π f (r ) r dr (D.6)
q 0
Z ∞
1 sin(qr )
fˇ(q) = 2 q dq f (q). (D.7)
2π r 0 r
Another formula which is often used is the Gaussian integral
∞ π b 2 +c
Z r
−ax 2 +bx+c
dx e = e 4a . (D.8)
−∞ a
Thus, consider the Fourier transform of (D.4) in spherical symmetry,

Z ∞ sin(qr )
n̂ ch (q) = 4π dr n ch (r )
0 q
X 4π A i Z ∞ h (r −Ri )2 (r +R i )2 i
= r d r sin(qr ) e γ2 + e γ2
3
i q 2π γ2 3 0
X A i ∂q Z ∞ h i
=− p 2 d u e i q(γu∓Ri ) + e −i q(γu∓Ri )
i 2q πγ −∞
q 2 γ2 h 2R i i
= A i e − 4 cos qR i +
X
sin qR i (D.9)
i qγ2
Now, it is reasonable to assume that the form factor of the proton is Gaussian-
shaped [57],
q e − (r −r 0 )2
F proton (r − r 0 ) = 3 e α2 , (D.10)
π 2 α3
where α is proportional to the root mean squared radius of the proton. The calcula-
tion of the Fourier transform of the shape factor is a simpler version of the previous
one; one finds
q 2 α2
F̂ proton (q) = q e e − 4 . (D.11)
We have all the ingredients necessary to use perform the deconvolution,
X A i − q 2 β2 h 2R i i
n̂(q) = e 4 cos qR i + 2
sin qR i , (D.12)
i qe qγ
91
where β2 = γ2 − α2 . The final step consists in a reverse Fourier transformation of the

nuclear density. Let us begin with the cosine term:
X Ai 1 Z ∞ sin(qr ) q 2 β2
− 4
n I (r ) = 2
q d q cos qR i e
i q e 2π r 0 r
Z ∞ i q 2 β2
Ai h
i q(r +R i ) −i q(r +R i ) i q(r −R i ) −i q(r −R i ) − 4
∂
X
= r d q e + e + e + e e
i 16q e π r
2
−∞
(r +R i )2 (r −R i )2
X Ai h −
β2
−
β2
i
= 3
(r + R i )e + (r − R i )e . (D.13)
i 2q e π 2 β3 r
As for the second term, a similar calculation gives
X A i Ri 1 Z ∞ sin(qr ) q 2 β2
n I I (r ) = q dq sin qR i e − 4
i q e γ 2π r 0
2 2 r
i qr
X A i Ri ∞ e − e −i qr e i qRi − e −i qRi − q 2 β2
Z
= d q e 4
i 2q e π γ r −∞
2 2 2i 2i
(r −R i )2 (r +R i )2
X A i Ri h −
β2
−
β2
i
= 3
e −e . (D.14)
i 2q e π γ2 βr
2
The final result reads

2 2
X Ai h³ r − R
i R i ´ − (r −R
β2
i) ³r +R
i R i ´ − (r +R
β2
i) i
n(r ) = + e + − e . (D.15)
i
3
2q e π 2 βr β2 γ2 β2 γ2
Appendix E
Discretization Methods
E.0.1 Discretized Derivative

The first and the second derivative are implemented through a symmetric five-
points formula, which is obtained by solving the following system of Taylor series
with respect to f (1) (x 0 ) and f (2) (x 0 ):
h 2 (2) h 3 (3) h 4 (4)

f (x 0 ± h) = f (x 0 ) ± h f (1) (x 0 ) + f (x 0 ) ± f (x 0 ) + f (x 0 ) + O(h 5 ), (E.1)
2 6 24
4h 2 (2) 8h 3 (3) 16h 4 (4)
f (x 0 ± 2h) = f (x 0 ) ± 2h f (1) (x 0 ) + f (x 0 ) ± f (x 0 ) + f (x 0 ) + O(h 5 ).
2 6 24
(E.2)
The system is solved by
− f (x 0 + 2h) + 8 f (x 0 + h) − 8 f (x 0 − h) + f (x 0 − 2h)
f (1) (x 0 ) = + O(h 4 ), (E.3)
12h
− f (x 0 + 2h) + 16 f (x 0 + h) − 30 f (x 0 ) + 16 f (x 0 − h) − f (x 0 − 2h)
f (2) (x 0 ) = + O(h 4 ).
12h 2
(E.4)
It is plain that these formulas are very advantageous, since they require only four
neighbours points, while the residuals wipe each other out and decay fast. On the
other hand, the formulas cannot be used blindly on the borders of a spatial grid.
The solution to this issue is to develop a non-symmetric derivative formula to be
used on the outer points. This choice has the advantage of being precise, but the
drawback of rendering the structure of the objective function, and of its gradient,
very involved; moreover, it causes the loss of the symmetry of the Hessian of the
Lagrangian. The asymmetric derivatives implementation can be found, at each order
of approximation, in many references, e. g. in [58].
E.0.2 Discretized Integration

Integrations have been implemented through two different methods. The first
single-particle test of this work uses a rectangles method, namely the Riemann
formula. This integration has a very low impact on the objective structure, but it is
93
94 APPENDIX E. DISCRETIZATION METHODS
very rough, and potentially brings along non-negligible numerical errors.

nX
r −1 h
Z
d x f (x) = h f (x i ) + ( f (x 1 ) + f (x nr )) + O(h) (E.5)
i =2 2
In order to improve the computation efficiency, we have implemented the Simp-

son’s integration method, which interpolates the integrand through second order
polynomials. This method is the one we have used in all of the test except the very
first. This formula is symmetric, and it is advantageous for those same reasons of the
symmetric derivative formula (E.3). Simpson’s quadrature formula reads
h³
Z ´
d x f (x) = f (x 0 ) + · · · + 4 f (x 2k ) + 2 f (x 2k+1 ) + · · · + f (x nr −1 ) + O(h 5 ). (E.6)
3
E.0.3 Linear Solver

Since the implementation of the discrete methods becomes more and more
complicated, the direct extraction of the Kohn-Sham potential from the Lagrange
multipliers becomes involved to be coded in a smart, general way. The correspon-
dence between the Lagrange multiplier and the Kohn-Sham potential gets partially
lost in the discretization of the space and of the numerical methods. Also, the com-
bined usage of two scaling procedures of the problem (one is the redefinition of the
orbitals according to the square root of the density, the second is an internal scaling
performed by the IPOPT library) contributes to mask the direct relation between the
numerical multipliers and the analytic ones. It is far better to separate the calculation
of the Kohn-Sham potential and of the multipliers ² j k from that of the Kohn-Sham
orbitals.
Once IPOPT gives us the Kohn-Sham orbitals, we use them to solve directly the
Euler-Lagrange equations with respect to the Lagrange multipliers. This is done via
a sparse least-squares linear solver, that we have coded using the EIGEN 1 library’s
QR decomposition method. The idea is to decompose the matrix A of the equations’
coefficients, which is actually a sparse one, as A = Q ∗ R, where Q is orthogonal, that
is straightforward to be inverted, and R is a sparse triangular or trapezoidal sparse
matrix. The solution of the system Ax = b is given by x = R −1Q −1 b = R −1Q t b. The
exploitation of the sparsity of the matrix improves the computational time.
E.0.4 Implementation
If we use the Riemann formula, together with the five-points derivatives, the
objective function takes the form
r −1 nX
1 nX p −1 ³ e 0 )2 n
(n e00
J {ϕ j (x)} ∼ − ϕi + j nr ϕi + j nr (− i + i )h
2 i =0 j =0 4n i 2
−ϕi + j nr +2 + 8ϕi + j nr +1 − 8ϕi + j nr −1 + ϕi + j nr −2
ei0
+n
12
−ϕi + j nr +2 + 16ϕi + j nr +1 − 30ϕi + j nr + 16ϕi + j nr −1 − ϕi + j nr −2 )
+n
ei
12h
−ϕi + j nr +2 + 16ϕi + j nr +1 − 30ϕi + j nr + 16ϕi + j nr −1 − ϕi + j nr −2 ´
− 2δ , (E.7)
12h
1 https://eigen.tuxfamily.org/
95
where n r is the number of points in the grid, while n p is the number of particles of
the system considered. The gradient of the objective reads
∂ f (ϕi + j nr )³ (ne 0 )2 n
e00
= − ϕl +mnr (− l + l )h
∂ϕl +mnr 4n l 2
−ϕl +mnr +2 + 16ϕl +mnr +1 − 30ϕl +mnr + 16ϕl +mnr −1 − ϕl +mnr −2 )
+nel
12h
−ϕl +mnr +2 + 16ϕl +mnr +1 − 30ϕl +mnr + 16ϕl +mnr −1 − ϕl +mnr −2 ´
+ 2δ , (E.8)
12h
where p = 0, . . . , n r − 1 and q = 0, . . . , n p − 1. The constraints are
n p −1
c i0 = ϕ2i + j nr = 1,
X
(E.9)
j =0
nX
r −1
c jk = ei ϕi + j nr ϕi +knr = δ j k ,
hn (E.10)
i =0
with j = 0, . . . , np − 1 and k = 0, . . . , j .
n p (n p +1)
The Jacobian of the constraints is a rectangular (n r + 2 × n r n p ) matrix, and
reads
2ϕmnr
 
..
.
 
 
 
 2ϕl +mnr 
  (E.11)
 .. 

 . 

 2ϕnr −1+mnr 
... ei (δi + j nr ,l +mnr + δi +knr ,l +mnr )
hn ...
Note that the matrix is always wider than taller, and therefore the system of Euler-
Lagrange equations is never under-determined. By further deriving the results ob-
tained right above, we obtain the Hessian of the Lagrangian as a nr × np square
matrix:  
..
.
 
 
.. ei0 )2 ei00
 
(n n 5nei +2δ
. + 2λ0i + 2λi i
 

 (− 4n i + 2 )h + 2h


 .. 4(nei +2δ) .. 

 . − 3h . 
 (E.12)
. ..
 .. nei +2δ

 12h
. 

 .. 

 λi j n
ei h . 

..
.
Bibliography
[1] H. De Vries, C.W. De Jager, and C. De Vries. Nuclear charge-density-distribution

parameters from elastic electron scattering. Atomic Data and Nuclear Data
Tables, 36(3):495 – 536, 1987.
[2] J. Zenihiro, H. Sakaguchi, T. Murakami, M. Yosoi, Y. Yasuda, S. Terashima, Y. Iwao,

H. Takeda, M. Itoh, H. P. Yoshida, and M. Uchida. Neutron density distributions
of 204,206,208 Pb deduced via proton elastic scattering at E p = 295 mev. Phys. Rev.
C, 82:044611, Oct 2010.
[3] J. Dobaczewski, H. Flocard, and J. Treiner. Hartree-fock-bogolyubov description

of nuclei near the neutron-drip line. Nuclear Physics A, 422(1):103 – 139, 1984.
[4] Richard M. Martin. Electronic Structure: Basic Theory and Practical Methods.
Cambridge University Press, 2004.
[5] X. Roca-Maza and N. Paar. Nuclear equation of state from ground and collective
excited state properties of nuclei. Progress in Particle and Nuclear Physics, 101:96
– 176, 2018.
[6] Michael Bender, Paul-Henri Heenen, and Paul-Gerhard Reinhard. Self-

consistent mean-field models for nuclear structure. Rev. Mod. Phys., 75:121–180,
Jan 2003.
[7] P.-G. Reinhard and W. Nazarewicz. Nuclear charge and neutron radii and nuclear
matter: Trend analysis in skyrme density-functional-theory approach. Phys. Rev.
C, 93:051303, May 2016.
[8] P. Hohenberg and W. Kohn. Inhomogeneous electron gas. Phys. Rev., 136:B864–
B871, Nov 1964.
[9] W. Kohn and L. J. Sham. Self-consistent equations including exchange and

correlation effects. Phys. Rev., 140:A1133–A1138, Nov 1965.
[10] R. Newton. Inverse problems in physics. SIAM Review, 12(3):346–356, 1970.
[11] Ashley H. Carter. A class of inverse problems in physics. American Journal of

Physics, 68(8):698–703, 2000.
[12] I. Macocco. Costruzione Del Potenziale Nucleare Efficace Di Kohn-Sham Dalle

Densità Empiriche. Bachelor Thesis. Not Published, 2015.
[13] A. Porro. Potenziali nucleari da densità e distribuzioni di spin microscopiche.

Bachelor Thesis. Not Published, 2017.
97
98 BIBLIOGRAPHY
[14] N. Martinez Cuenca. Deduzione di potenziale medio e massa efficace dei nucleoni
dalle densita nucleari. Bachelor Thesis. Not Published, 2018.
[15] Weitao Yang and Qin Wu. Direct method for optimized effective potentials in
density-functional theory. Phys. Rev. Lett., 89:143002, Sep 2002.
[16] Qin Wu and Weitao Yang. A direct optimization method for calculating density
functionals and exchange–correlation potentials from electron densities. The
Journal of Chemical Physics, 118(6):2498–2509, 2003.
[17] Qingsheng Zhao, Robert C. Morrison, and Robert G. Parr. From electron densities
to kohn-sham kinetic energies, orbital energies, exchange-correlation potentials,
and exchange-correlation energies. Phys. Rev. A, 50:2138–2142, Sep 1994.
[18] R. van Leeuwen and E. J. Baerends. Exchange-correlation potential with correct

asymptotic behavior. Phys. Rev. A, 49:2421–2431, Apr 1994.
[19] Yue Wang and Robert G. Parr. Construction of exact kohn-sham orbitals from a
given electron density. Phys. Rev. A, 47:R1591–R1593, Mar 1993.
[20] Qingsheng Zhao and Robert G. Parr. Quantities ts [n] and tc [n] in density-
functional theory. Phys. Rev. A, 46:2337–2343, Sep 1992.
[21] Andreas Görling. Kohn-sham potentials and wave functions from electron
densities. Phys. Rev. A, 46:3753–3757, Oct 1992.
[22] Daniel S. Jensen and Adam Wasserman. Numerical density-to-potential inver-

sions in time-dependent density functional theory. Phys. Chem. Chem. Phys.,
18:21079–21091, 2016.
[23] M.G. Mayer and J.H.D. Jensen. Elementary Theory of Nuclear Shell Structure.
Structure of matter series. John Wiley & Sons, 1955.
[24] A. L. Fetter and J. D. Walecka. Quantum Theory of Many-Particle Systems (Dover

Books on Physics). Dover Publications, June 2003.
[25] P. Ring and P. Schuck. The Nuclear Many-Body Problem (Theoretical and Mathe-
matical Physics). Springer, May 2005.
[26] A. Bohr and B.R. Mottelson. Nuclear Structure. Number v. 1 in Nuclear Structure.
World Scientific, 1998.
[27] T. H. R. Skyrme. The effective nuclear potential. Nuclear Physics, 9(4):615 – 634,
1958–1959.
[28] J. Dechargé and D. Gogny. Hartree-fock-bogolyubov calculations with the d 1

effective interaction on spherical nuclei. Phys. Rev. C, 21:1568–1593, Apr 1980.
[29] T. Nikšić, D. Vretenar, and P. Ring. Relativistic nuclear energy density functionals:
Mean-field and beyond. Progress in Particle and Nuclear Physics, 66(3):519 – 548,
2011.
[30] W. Kohn. v-representability and density functional theory. Phys. Rev. Lett.,
51:1596–1598, Oct 1983.
BIBLIOGRAPHY 99
[31] H. Englisch and R. Englisch. Hohenberg-kohn theorem and non-v-representable

densities. Physica A: Statistical Mechanics and its Applications, 121(1):253 – 268,
1983.
[32] T. L. Gilbert. Hohenberg-kohn theorem for nonlocal external potentials. Phys.

Rev. B, 12:2111–2120, Sep 1975.
[33] R.G. Parr and Y. Weitao. Density Functional Theory of Atoms and Molecules.
International Series of Monographs on Chemistry. Oxford University Press, 1994.
[34] Mel Levy. Electron densities in search of hamiltonians. Phys. Rev. A, 26:1200–
1208, Sep 1982.
[35] Elliott H. Lieb. Density Functionals for Coulomb Systems, pages 269–303. Springer
Berlin Heidelberg, Berlin, Heidelberg, 2002.
[36] Klaus Capelle. A bird’s-eye view of density-functional theory. Brazilian Journal

of Physics, 36(4A):1318–1343, 2006.
[37] J. Kaipio and E. Somersalo. Statistical and Computational Inverse Problems.

Applied Mathematical Sciences. Springer New York, 2006.
[38] V. I. Arnold. Catastrophe Theory. Springer-Verlag Berlin Heidelberg, 1984.
[39] Jari Kaipio and Erkki Somersalo. Statistical inverse problems: Discretization,
model reduction and inverse crimes. Journal of Computational and Applied
Mathematics, 198(2):493 – 504, 2007. Special Issue: Applied Computational
Inverse Problems.
[40] A. N. Tikhonov. On the solution of ill-posed problems and the method of regu-
larization. Dokl. Akad. Nauk SSSR, 151:501–504, 1963.
[41] O. Scherzer. The use of morozov’s discrepancy principle for tikhonov regular-
ization for solving nonlinear ill-posed problems. Computing, 51(1):45–60, Mar
1993.
[42] J. T. Chayes, L. Chayes, and Mary Beth Ruskai. Density functional approach to
quantum lattice systems. Journal of Statistical Physics, 38(3):497–518, Feb 1985.
[43] M. E. Rose. The charge distribution in nuclei and the scattering of high energy
electrons. Phys. Rev, 73:279 – 284, Feb 1948.
[44] Robert Hofstadter. Electron scattering and nuclear structure. Rev. Mod. Phys.,
28:214–254, Jul 1956.
[45] J. Negele. Advances in Nuclear Physics. Number v. 15 in Advances in Nuclear

Physics. Springer US, 2013.
[46] Andreas Wächter and Lorenz T. Biegler. On the implementation of an interior-

point filter line-search algorithm for large-scale nonlinear programming. Math-
ematical Programming, 106(1):25–57, Mar 2006.
[47] Philip M. Morse. Diatomic molecules according to the wave mechanics. ii.
vibrational levels. Phys. Rev., 34:57–64, Jul 1929.
100 BIBLIOGRAPHY
[48] S. Wilson. Handbook of Molecular Physics and Quantum Chemistry: Molecules

in the physico-chemical environment: spectroscopy, dynamics and bulk proper-
ties. Handbook of Molecular Physics and Quantum Chemistry: Spectroscopy,
Dynamics and Bulk Properties. Molecules in the Physico-chemical Environment.
Wiley, 2003.
[49] Jens Peter Dahl and Michael Springborg. The morse oscillator in positition space,
momentum space, and phase space. Journal of Chemichal Pysics, 88(7):4535–
4547, 1988.
[50] J. Lafontaine. An Introduction to Differential Manifolds. Springer International

Publishing, 2015.
[51] D.P. Bertsekas and W. Rheinboldt. Constrained Optimization and Lagrange Mul-
tiplier Methods. Computer science and applied mathematics. Elsevier Science,
2014.
[52] H. P. Breuer and F. Petruccione. The theory of open quantum systems. Oxford
University Press, Great Clarendon Street, 2002.
[53] J. Dobaczewski and J. Dudek. Time-odd components in the rotating mean field
and identical bands. Acta Phys.Polon. (1996), B27:45–58, Oct 1995.
[54] J. Dobaczewski and J. Dudek. Time-odd components in the mean field of

rotating superdeformed nuclei. Phys. Rev. C, 52:1827–1839, Jan 1995.
[55] K. Heyde. Basic Ideas and Concepts in Nuclear Physics: An Introductory Approach,
Second Edition. Taylor & Francis, 1998.
[56] D.S. Koltun and J.M. Eisenberg. Quantum mechanics of many degrees of freedom.
Wiley-Interscience publication. Wiley, 1988.
[57] Zhihong Ye, John Arrington, Richard J. Hill, and Gabriel Lee. Proton and neutron
electromagnetic form factors and uncertainties. Physics Letters B, 777:8 – 15,
2018.
[58] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. Numerical Recipes

3rd Edition: The Art of Scientific Computing. Cambridge University Press, New
York, NY, USA, 3 edition, 2007.

Thesis Giacomo Accorto Master

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis Giacomo Accorto Master

Uploaded by

Copyright:

Available Formats

Università degli Studi di Milano

FACOLTÀ DI SCIENZE E TECNOLOGIE

Tesi di Laurea Magistrale

Anno Accademico 2017–2018

2 IKS Problem in Nuclear DFT 19

3 Implementation and Inversion Results 43

A The Method of Lagrange Multipliers 81

B Density Operators and Scalar/Vector Densities 83

D Deconvolution from Proton Charge Density to Proton Density 89

Nuclear Density Functional

1.1 Introduction to Nuclear Phenomenology and to Ba-

1.1.1 Nuclear forces and Bulk Properties of Nuclei

n(0) = 0.16 fm−3 (1.1)

However, it is a noteworthy example of how different effects interplay in nuclear

1.1.2 Basic Concepts of Quantum Mechanics

1.1.3 Nuclear Shell Model

H 0 (r ) = −α(r )L̂ · Ŝ, (1.25)

1.2 Density Functional Theory

and it integrates to the total number of nucleons in the system

1.2.1 Hohenberg-Kohn Theorems

Proof. The density trivially determines A by quadrature. It is then just necessary to

E 0 < 〈Ψ0 | Ĥ |Ψ0 〉 = 〈Ψ0 | Ĥ 0 |Ψ0 〉 + 〈Ψ0 | Ĥ − Ĥ 0 |Ψ0 〉

The same is indeed true if we interchange the dummy primed variables

E 00 < 〈Ψ| Ĥ 0 |Ψ〉 = 〈Ψ| Ĥ |Ψ〉 + 〈Ψ0 | Ĥ 0 − Ĥ |Ψ0 〉

The ground state energy is a global minumum of the total energy,

Proof. Because of the first Hohenberg and Kohn theorem, n

If E v [n] is assumed to be differentiable, the third Hohenberg theorem can be

where µ is the chemical potential

1.3 Levy-Lieb Constrained-Search Formulation of DFT

1 Except for some simple dicrete systems [30]

where k = 0, ±1, ±2, . . . or k = ± 12 , ± 32 , . . . ; it holds that

Any given one-dimensional density can represented as

1.3.2 Levy-Lieb Constrained-Search Formulation

〈Ψn0 | Ĥ |Ψn0 〉 ≥ 〈Ψ0 | Ĥ |Ψ0 〉 = E 0 . (1.49)

F H K [n 0 ] = 〈Ψ0 |T̂ + Ŵ |Ψ0 〉 , (1.50)

F LL [n] = min 〈Ψ|T̂ + Ŵ |Ψ〉 . (1.52)

E 0 = min 〈Ψ|T̂ + Ŵ + V̂ |Ψ〉 (1.53)

1.4 Kohn-Sham Density Functional Theory

F [n] = TK S [n] +U [n] + E xc [n], (1.63)

and the exchange-correlation potential is

Among the state-of-art beyond-LDA approximation schemes the most relevant is

Inverse Problem in Nuclear

2.1 Forward and Inverse Problems

2.2 Regularization techniques for Ill-posed Problems

• the existence of a solution of the problem;

• the uniqueness of the solution;

2.2.1 Tikhonov Regularization

F δ (x) = kAx − yk2 + δkxk2 , (2.1)

provided that such minimum exists. The regularization parameter represents a

can be considered acceptable as an approximate solution. If x δ is the regularized

f δ = kAx δ − yk. (2.3)

2.3 The Inverse Problem In Nuclear Kohn-Sham Den-

2.3.2 Inversion methods

advantages and drawbacks. We then promote and suggest a combined usage of

2.4 Constrained Variational Inversion Method

where j = 1, . . . , N and k = j , . . . , N are an adequate set of quantum numbers. The

the Kohn-Sham orbitals to be real. In the appendix C a formal justification of this

The derivation of the constraint terms gives

The sum of the two equations produces

Again, the solution of the Euler-Lagrange equations

optimizes the functional, and an explicit calculation of those gives

Once more, the artificial introduction of g β (~

~2 ³ (∇n(~e r ))2 ∇2 n(~

2.5 Insertion of Spherical Symmetry into the Constrained